13 skills found
molyswu / Hand Detectionusing Neural Networks (SSD) on Tensorflow. This repo documents steps and scripts used to train a hand detector using Tensorflow (Object Detection API). As with any DNN based task, the most expensive (and riskiest) part of the process has to do with finding or creating the right (annotated) dataset. I was interested mainly in detecting hands on a table (egocentric view point). I experimented first with the [Oxford Hands Dataset](http://www.robots.ox.ac.uk/~vgg/data/hands/) (the results were not good). I then tried the [Egohands Dataset](http://vision.soic.indiana.edu/projects/egohands/) which was a much better fit to my requirements. The goal of this repo/post is to demonstrate how neural networks can be applied to the (hard) problem of tracking hands (egocentric and other views). Better still, provide code that can be adapted to other uses cases. If you use this tutorial or models in your research or project, please cite [this](#citing-this-tutorial). Here is the detector in action. <img src="images/hand1.gif" width="33.3%"><img src="images/hand2.gif" width="33.3%"><img src="images/hand3.gif" width="33.3%"> Realtime detection on video stream from a webcam . <img src="images/chess1.gif" width="33.3%"><img src="images/chess2.gif" width="33.3%"><img src="images/chess3.gif" width="33.3%"> Detection on a Youtube video. Both examples above were run on a macbook pro **CPU** (i7, 2.5GHz, 16GB). Some fps numbers are: | FPS | Image Size | Device| Comments| | ------------- | ------------- | ------------- | ------------- | | 21 | 320 * 240 | Macbook pro (i7, 2.5GHz, 16GB) | Run without visualizing results| | 16 | 320 * 240 | Macbook pro (i7, 2.5GHz, 16GB) | Run while visualizing results (image above) | | 11 | 640 * 480 | Macbook pro (i7, 2.5GHz, 16GB) | Run while visualizing results (image above) | > Note: The code in this repo is written and tested with Tensorflow `1.4.0-rc0`. Using a different version may result in [some errors](https://github.com/tensorflow/models/issues/1581). You may need to [generate your own frozen model](https://pythonprogramming.net/testing-custom-object-detector-tensorflow-object-detection-api-tutorial/?completed=/training-custom-objects-tensorflow-object-detection-api-tutorial/) graph using the [model checkpoints](model-checkpoint) in the repo to fit your TF version. **Content of this document** - Motivation - Why Track/Detect hands with Neural Networks - Data preparation and network training in Tensorflow (Dataset, Import, Training) - Training the hand detection Model - Using the Detector to Detect/Track hands - Thoughts on Optimizations. > P.S if you are using or have used the models provided here, feel free to reach out on twitter ([@vykthur](https://twitter.com/vykthur)) and share your work! ## Motivation - Why Track/Detect hands with Neural Networks? There are several existing approaches to tracking hands in the computer vision domain. Incidentally, many of these approaches are rule based (e.g extracting background based on texture and boundary features, distinguishing between hands and background using color histograms and HOG classifiers,) making them not very robust. For example, these algorithms might get confused if the background is unusual or in situations where sharp changes in lighting conditions cause sharp changes in skin color or the tracked object becomes occluded.(see [here for a review](https://www.cse.unr.edu/~bebis/handposerev.pdf) paper on hand pose estimation from the HCI perspective) With sufficiently large datasets, neural networks provide opportunity to train models that perform well and address challenges of existing object tracking/detection algorithms - varied/poor lighting, noisy environments, diverse viewpoints and even occlusion. The main drawbacks to usage for real-time tracking/detection is that they can be complex, are relatively slow compared to tracking-only algorithms and it can be quite expensive to assemble a good dataset. But things are changing with advances in fast neural networks. Furthermore, this entire area of work has been made more approachable by deep learning frameworks (such as the tensorflow object detection api) that simplify the process of training a model for custom object detection. More importantly, the advent of fast neural network models like ssd, faster r-cnn, rfcn (see [here](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md#coco-trained-models-coco-models) ) etc make neural networks an attractive candidate for real-time detection (and tracking) applications. Hopefully, this repo demonstrates this. > If you are not interested in the process of training the detector, you can skip straight to applying the [pretrained model I provide in detecting hands](#detecting-hands). Training a model is a multi-stage process (assembling dataset, cleaning, splitting into training/test partitions and generating an inference graph). While I lightly touch on the details of these parts, there are a few other tutorials cover training a custom object detector using the tensorflow object detection api in more detail[ see [here](https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/) and [here](https://towardsdatascience.com/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9) ]. I recommend you walk through those if interested in training a custom object detector from scratch. ## Data preparation and network training in Tensorflow (Dataset, Import, Training) **The Egohands Dataset** The hand detector model is built using data from the [Egohands Dataset](http://vision.soic.indiana.edu/projects/egohands/) dataset. This dataset works well for several reasons. It contains high quality, pixel level annotations (>15000 ground truth labels) where hands are located across 4800 images. All images are captured from an egocentric view (Google glass) across 48 different environments (indoor, outdoor) and activities (playing cards, chess, jenga, solving puzzles etc). <img src="images/egohandstrain.jpg" width="100%"> If you will be using the Egohands dataset, you can cite them as follows: > Bambach, Sven, et al. "Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions." Proceedings of the IEEE International Conference on Computer Vision. 2015. The Egohands dataset (zip file with labelled data) contains 48 folders of locations where video data was collected (100 images per folder). ``` -- LOCATION_X -- frame_1.jpg -- frame_2.jpg ... -- frame_100.jpg -- polygons.mat // contains annotations for all 100 images in current folder -- LOCATION_Y -- frame_1.jpg -- frame_2.jpg ... -- frame_100.jpg -- polygons.mat // contains annotations for all 100 images in current folder ``` **Converting data to Tensorflow Format** Some initial work needs to be done to the Egohands dataset to transform it into the format (`tfrecord`) which Tensorflow needs to train a model. This repo contains `egohands_dataset_clean.py` a script that will help you generate these csv files. - Downloads the egohands datasets - Renames all files to include their directory names to ensure each filename is unique - Splits the dataset into train (80%), test (10%) and eval (10%) folders. - Reads in `polygons.mat` for each folder, generates bounding boxes and visualizes them to ensure correctness (see image above). - Once the script is done running, you should have an images folder containing three folders - train, test and eval. Each of these folders should also contain a csv label document each - `train_labels.csv`, `test_labels.csv` that can be used to generate `tfrecords` Note: While the egohands dataset provides four separate labels for hands (own left, own right, other left, and other right), for my purpose, I am only interested in the general `hand` class and label all training data as `hand`. You can modify the data prep script to generate `tfrecords` that support 4 labels. Next: convert your dataset + csv files to tfrecords. A helpful guide on this can be found [here](https://pythonprogramming.net/creating-tfrecord-files-tensorflow-object-detection-api-tutorial/).For each folder, you should be able to generate `train.record`, `test.record` required in the training process. ## Training the hand detection Model Now that the dataset has been assembled (and your tfrecords), the next task is to train a model based on this. With neural networks, it is possible to use a process called [transfer learning](https://www.tensorflow.org/tutorials/image_retraining) to shorten the amount of time needed to train the entire model. This means we can take an existing model (that has been trained well on a related domain (here image classification) and retrain its final layer(s) to detect hands for us. Sweet!. Given that neural networks sometimes have thousands or millions of parameters that can take weeks or months to train, transfer learning helps shorten training time to possibly hours. Tensorflow does offer a few models (in the tensorflow [model zoo](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md#coco-trained-models-coco-models)) and I chose to use the `ssd_mobilenet_v1_coco` model as my start point given it is currently (one of) the fastest models (read the SSD research [paper here](https://arxiv.org/pdf/1512.02325.pdf)). The training process can be done locally on your CPU machine which may take a while or better on a (cloud) GPU machine (which is what I did). For reference, training on my macbook pro (tensorflow compiled from source to take advantage of the mac's cpu architecture) the maximum speed I got was 5 seconds per step as opposed to the ~0.5 seconds per step I got with a GPU. For reference it would take about 12 days to run 200k steps on my mac (i7, 2.5GHz, 16GB) compared to ~5hrs on a GPU. > **Training on your own images**: Please use the [guide provided by Harrison from pythonprogramming](https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/) on how to generate tfrecords given your label csv files and your images. The guide also covers how to start the training process if training locally. [see [here] (https://pythonprogramming.net/training-custom-objects-tensorflow-object-detection-api-tutorial/)]. If training in the cloud using a service like GCP, see the [guide here](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_on_cloud.md). As the training process progresses, the expectation is that total loss (errors) gets reduced to its possible minimum (about a value of 1 or thereabout). By observing the tensorboard graphs for total loss(see image below), it should be possible to get an idea of when the training process is complete (total loss does not decrease with further iterations/steps). I ran my training job for 200k steps (took about 5 hours) and stopped at a total Loss (errors) value of 2.575.(In retrospect, I could have stopped the training at about 50k steps and gotten a similar total loss value). With tensorflow, you can also run an evaluation concurrently that assesses your model to see how well it performs on the test data. A commonly used metric for performance is mean average precision (mAP) which is single number used to summarize the area under the precision-recall curve. mAP is a measure of how well the model generates a bounding box that has at least a 50% overlap with the ground truth bounding box in our test dataset. For the hand detector trained here, the mAP value was **0.9686@0.5IOU**. mAP values range from 0-1, the higher the better. <img src="images/accuracy.jpg" width="100%"> Once training is completed, the trained inference graph (`frozen_inference_graph.pb`) is then exported (see the earlier referenced guides for how to do this) and saved in the `hand_inference_graph` folder. Now its time to do some interesting detection. ## Using the Detector to Detect/Track hands If you have not done this yet, please following the guide on installing [Tensorflow and the Tensorflow object detection api](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md). This will walk you through setting up the tensorflow framework, cloning the tensorflow github repo and a guide on - Load the `frozen_inference_graph.pb` trained on the hands dataset as well as the corresponding label map. In this repo, this is done in the `utils/detector_utils.py` script by the `load_inference_graph` method. ```python detection_graph = tf.Graph() with detection_graph.as_default(): od_graph_def = tf.GraphDef() with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid: serialized_graph = fid.read() od_graph_def.ParseFromString(serialized_graph) tf.import_graph_def(od_graph_def, name='') sess = tf.Session(graph=detection_graph) print("> ====== Hand Inference graph loaded.") ``` - Detect hands. In this repo, this is done in the `utils/detector_utils.py` script by the `detect_objects` method. ```python (boxes, scores, classes, num) = sess.run( [detection_boxes, detection_scores, detection_classes, num_detections], feed_dict={image_tensor: image_np_expanded}) ``` - Visualize detected bounding detection_boxes. In this repo, this is done in the `utils/detector_utils.py` script by the `draw_box_on_image` method. This repo contains two scripts that tie all these steps together. - detect_multi_threaded.py : A threaded implementation for reading camera video input detection and detecting. Takes a set of command line flags to set parameters such as `--display` (visualize detections), image parameters `--width` and `--height`, videe `--source` (0 for camera) etc. - detect_single_threaded.py : Same as above, but single threaded. This script works for video files by setting the video source parameter videe `--source` (path to a video file). ```cmd # load and run detection on video at path "videos/chess.mov" python detect_single_threaded.py --source videos/chess.mov ``` > Update: If you do have errors loading the frozen inference graph in this repo, feel free to generate a new graph that fits your TF version from the model-checkpoint in this repo. Use the [export_inference_graph.py](https://github.com/tensorflow/models/blob/master/research/object_detection/export_inference_graph.py) script provided in the tensorflow object detection api repo. More guidance on this [here](https://pythonprogramming.net/testing-custom-object-detector-tensorflow-object-detection-api-tutorial/?completed=/training-custom-objects-tensorflow-object-detection-api-tutorial/). ## Thoughts on Optimization. A few things that led to noticeable performance increases. - Threading: Turns out that reading images from a webcam is a heavy I/O event and if run on the main application thread can slow down the program. I implemented some good ideas from [Adrian Rosebuck](https://www.pyimagesearch.com/2017/02/06/faster-video-file-fps-with-cv2-videocapture-and-opencv/) on parrallelizing image capture across multiple worker threads. This mostly led to an FPS increase of about 5 points. - For those new to Opencv, images from the `cv2.read()` method return images in [BGR format](https://www.learnopencv.com/why-does-opencv-use-bgr-color-format/). Ensure you convert to RGB before detection (accuracy will be much reduced if you dont). ```python cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB) ``` - Keeping your input image small will increase fps without any significant accuracy drop.(I used about 320 x 240 compared to the 1280 x 720 which my webcam provides). - Model Quantization. Moving from the current 32 bit to 8 bit can achieve up to 4x reduction in memory required to load and store models. One way to further speed up this model is to explore the use of [8-bit fixed point quantization](https://heartbeat.fritz.ai/8-bit-quantization-and-tensorflow-lite-speeding-up-mobile-inference-with-low-precision-a882dfcafbbd). Performance can also be increased by a clever combination of tracking algorithms with the already decent detection and this is something I am still experimenting with. Have ideas for optimizing better, please share! <img src="images/general.jpg" width="100%"> Note: The detector does reflect some limitations associated with the training set. This includes non-egocentric viewpoints, very noisy backgrounds (e.g in a sea of hands) and sometimes skin tone. There is opportunity to improve these with additional data. ## Integrating Multiple DNNs. One way to make things more interesting is to integrate our new knowledge of where "hands" are with other detectors trained to recognize other objects. Unfortunately, while our hand detector can in fact detect hands, it cannot detect other objects (a factor or how it is trained). To create a detector that classifies multiple different objects would mean a long involved process of assembling datasets for each class and a lengthy training process. > Given the above, a potential strategy is to explore structures that allow us **efficiently** interleave output form multiple pretrained models for various object classes and have them detect multiple objects on a single image. An example of this is with my primary use case where I am interested in understanding the position of objects on a table with respect to hands on same table. I am currently doing some work on a threaded application that loads multiple detectors and outputs bounding boxes on a single image. More on this soon.
abusufyanvu / 6S191 MIT DeepLearningMIT Introduction to Deep Learning (6.S191) Instructors: Alexander Amini and Ava Soleimany Course Information Summary Prerequisites Schedule Lectures Labs, Final Projects, Grading, and Prizes Software labs Gather.Town lab + Office Hour sessions Final project Paper Review Project Proposal Presentation Project Proposal Grading Rubric Past Project Proposal Ideas Awards + Categories Important Links and Emails Course Information Summary MIT's introductory course on deep learning methods with applications to computer vision, natural language processing, biology, and more! Students will gain foundational knowledge of deep learning algorithms and get practical experience in building neural networks in TensorFlow. Course concludes with a project proposal competition with feedback from staff and a panel of industry sponsors. Prerequisites We expect basic knowledge of calculus (e.g., taking derivatives), linear algebra (e.g., matrix multiplication), and probability (e.g., Bayes theorem) -- we'll try to explain everything else along the way! Experience in Python is helpful but not necessary. This class is taught during MIT's IAP term by current MIT PhD researchers. Listeners are welcome! Schedule Monday Jan 18, 2021 Lecture: Introduction to Deep Learning and NNs Lab: Lab 1A Tensorflow and building NNs from scratch Tuesday Jan 19, 2021 Lecture: Deep Sequence Modelling Lab: Lab 1B Music Generation using RNNs Wednesday Jan 20, 2021 Lecture: Deep Computer Vision Lab: Lab 2A Image classification and detection Thursday Jan 21, 2021 Lecture: Deep Generative Modelling Lab: Lab 2B Debiasing facial recognition systems Friday Jan 22, 2021 Lecture: Deep Reinforcement Learning Lab: Lab 3 pixel-to-control planning Monday Jan 25, 2021 Lecture: Limitations and New Frontiers Lab: Lab 3 continued Tuesday Jan 26, 2021 Lecture (part 1): Evidential Deep Learning Lecture (part 2): Bias and Fairness Lab: Work on final assignments Lab competition entries due at 11:59pm ET on Canvas! Lab 1, Lab 2, and Lab 3 Wednesday Jan 27, 2021 Lecture (part 1): Nigel Duffy, Ernst & Young Lecture (part 2): Kate Saenko, Boston University and MIT-IBM Watson AI Lab Lab: Work on final assignments Assignments due: Sign up for Final Project Competition Thursday Jan 28, 2021 Lecture (part 1): Sanja Fidler, U. Toronto, Vector Institute, and NVIDIA Lecture (part 2): Katherine Chou, Google Lab: Work on final assignments Assignments due: 1 page paper review (if applicable) Friday Jan 29, 2021 Lecture: Student project pitch competition Lab: Awards ceremony and prize giveaway Assignments due: Project proposals (if applicable) Lectures Lectures will be held starting at 1:00pm ET from Jan 18 - Jan 29 2021, Monday through Friday, virtually through Zoom. Current MIT students, faculty, postdocs, researchers, staff, etc. will be able to access the lectures during this two week period, synchronously or asynchronously, via the MIT Canvas course webpage (MIT internal only). Lecture recordings will be uploaded to the Canvas as soon as possible; students are not required to attend any lectures synchronously. Please see the Canvas for details on Zoom links. The public edition of the course will only be made available after completion of the MIT course. Labs, Final Projects, Grading, and Prizes Course will be graded during MIT IAP for 6 units under P/D/F grading. Receiving a passing grade requires completion of each software lab project (through honor code, with submission required to enter lab competitions), a final project proposal/presentation or written review of a deep learning paper (submission required), and attendance/lecture viewing (through honor code). Submission of a written report or presentation of a project proposal will ensure a passing grade. MIT students will be eligible for prizes and awards as part of the class competitions. There will be two parts to the competitions: (1) software labs and (2) final projects. More information is provided below. Winners will be announced on the last day of class, with thousands of dollars of prizes being given away! Software labs There are three TensorFlow software lab exercises for the course, designed as iPython notebooks hosted in Google Colab. Software labs can be found on GitHub: https://github.com/aamini/introtodeeplearning. These are self-paced exercises and are designed to help you gain practical experience implementing neural networks in TensorFlow. For registered MIT students, submission of lab materials is not necessary to get credit for the course or to pass the course. At the end of each software lab there will be task-associated materials to submit (along with instructions) for entry into the competitions, open to MIT students and affiliates during the IAP offering. This includes MIT students/affiliates who are taking the class as listeners -- you are eligible! These instructions are provided at the end of each of the labs. Completing these tasks and submitting your materials to Canvas will enter you into a per-lab competition. MIT students and affiliates will be eligible for prizes during the IAP offering; at the end of the course, prize-winners will be awarded with their prizes. All competition submissions are due on January 26 at 11:59pm ET to Canvas. For the software lab competitions, submissions will be judged on the basis of the following criteria: Strength and quality of final results (lab dependent) Soundness of implementation and approach Thoroughness and quality of provided descriptions and figures Gather.Town lab + Office Hour sessions After each day’s lecture, there will be open Office Hours in the class GatherTown, up until 3pm ET. An MIT email is required to log in and join the GatherTown. During these sessions, there will not be a walk through or dictation of the labs; the labs are designed to be self-paced and to be worked on on your own time. The GatherTown sessions will be hosted by course staff and are held so you can: Ask questions on course lectures, labs, logistics, project, or anything else; Work on the labs in the presence of classmates/TAs/instructors; Meet classmates to find groups for the final project; Group work time for the final project; Bring the class community together. Final project To satisfy the final project requirement for this course, students will have two options: (1) write a 1 page paper review (single-spaced) on a recent deep learning paper of your choice or (2) participate and present in the project proposal pitch competition. The 1 page paper review option is straightforward, we propose some papers within this document to help you get started, and you can satisfy a passing grade with this option -- you will not be eligible for the grand prizes. On the other hand, participation in the project proposal pitch competition will equivalently satisfy your course requirements but additionally make you eligible for the grand prizes. See the section below for more details and requirements for each of these options. Paper Review Students may satisfy the final project requirement by reading and reviewing a recent deep learning paper of their choosing. In the written review, students should provide both: 1) a description of the problem, technical approach, and results of the paper; 2) critical analysis and exposition of the limitations of the work and opportunities for future work. Reviews should be submitted on Canvas by Thursday Jan 28, 2021, 11:59:59pm Eastern Time (ET). Just a few paper options to consider... https://papers.nips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf https://papers.nips.cc/paper/2018/file/69386f6bb1dfed68692a24c8686939b9-Paper.pdf https://papers.nips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf https://science.sciencemag.org/content/362/6419/1140 https://papers.nips.cc/paper/2018/file/0e64a7b00c83e3d22ce6b3acf2c582b6-Paper.pdf https://arxiv.org/pdf/1906.11829.pdf https://www.nature.com/articles/s42256-020-00237-3 https://pubmed.ncbi.nlm.nih.gov/32084340/ Project Proposal Presentation Keyword: proposal This is a 2 week course so we do not require results or working implementations! However, to win the top prizes, nice, clear results and implementations will demonstrate feasibility of your proposal which is something we look for! Logistics -- please read! You must sign up to present before 11:59:59pm Eastern Time (ET) on Wednesday Jan 27, 2021 Slides must be in a Google Slide before 11:59:59pm Eastern Time (ET) on Thursday Jan 28, 2021 Project groups can be between 1 and 5 people Listeners welcome To be eligible for a prize you must have at least 1 registered MIT student in your group Each participant will only be allowed to be in one group and present one project pitch Synchronous attendance on 1/29/21 is required to make the project pitch! 3 min presentation on your idea (we will be very strict with the time limits) Prizes! (see below) Sign up to Present here: by 11:59pm ET on Wednesday Jan 27 Once you sign up, make your slide in the following Google Slides; submit by midnight on Thursday Jan 28. Please specify the project group # on your slides!!! Things to Consider This doesn’t have to be a new deep learning method. It can just be an interesting application that you apply some existing deep learning method to. What problem are you solving? Are there use cases/applications? Why do you think deep learning methods might be suited to this task? How have people done it before? Is it a new task? If so, what are similar tasks that people have worked on? In what aspects have they succeeded or failed? What is your method of solving this problem? What type of model + architecture would you use? Why? What is the data for this task? Do you need to make a dataset or is there one publicly available? What are the characteristics of the data? Is it sparse, messy, imbalanced? How would you deal with that? Project Proposal Grading Rubric Project proposals will be evaluated by a panel of judges on the basis of the following three criteria: 1) novelty and impact; 2) technical soundness, feasibility, and organization, including quality of any presented results; 3) clarity and presentation. Each judge will award a score from 1 (lowest) to 5 (highest) for each of the criteria; the average score from each judge across these criteria will then be averaged with that of the other judges to provide the final score. The proposals with the highest final scores will be selected for prizes. Here are the guidelines for the criteria: Novelty and impact: encompasses the potential impact of the project idea, its novelty with respect to existing approaches. Why does the proposed work matter? What problem(s) does it solve? Why are these problems important? Technical soundness, feasibility, and organization: encompasses all technical aspects of the proposal. Do the proposed methodology and architecture make sense? Is the architecture the best suited for the proposed problem? Is deep learning the best approach for the problem? How realistic is it to implement the idea? Was there any implementation of the method? If results and data are presented, we will evaluate the strength of the results/data. Clarity and presentation: encompasses the delivery and quality of the presentation itself. Is the talk well organized? Are the slides aesthetically compelling? Is there a clear, well-delivered narrative? Are the problem and proposed method clearly presented? Past Project Proposal Ideas Recipe Generation with RNNs Can we compress videos with CNN + RNN? Music Generation with RNNs Style Transfer Applied to X GAN’s on a new modality Summarizing text/news articles Combining news articles about similar events Code or spec generation Multimodal speech → handwriting Generate handwriting based on keywords (i.e. cursive, slanted, neat) Predicting stock market trends Show language learners articles or videos at their level Transfer of writing style Chemical Synthesis with Recurrent Neural networks Transfer learning to learn something in a domain for which it’s hard or risky to gather data or do training RNNs to model some type of time series data Computer vision to coach sports players Computer vision system for safety brakes or warnings Use IBM Watson API to get the sentiment of your Facebook newsfeed Deep learning webcam to give wifi-access to friends or improve video chat in some way Domain-specific chatbot to help you perform a specific task Detect whether a signature is fraudulent Awards + Categories Final Project Awards: 1x NVIDIA RTX 3080 4x Google Home Max 3x Display Monitors Software Lab Awards: Bose headphones (Lab 1) Display monitor (Lab 2) Bebop drone (Lab 3) Important Links and Emails Course website: http://introtodeeplearning.com Course staff: introtodeeplearning-staff@mit.edu Piazza forum (MIT only): https://piazza.com/mit/spring2021/6s191 Canvas (MIT only): https://canvas.mit.edu/courses/8291 Software lab repository: https://github.com/aamini/introtodeeplearning Lab/office hour sessions (MIT only): https://gather.town/app/56toTnlBrsKCyFgj/MITDeepLearning
ShanJiangEmugen / Conv3d Video ToolkitA lightweight and modular 3D ConvNet pipeline for video-based behavior classification. Includes data preparation tools, frame-level video generators, Conv3D training scripts, fine-tuning utilities, and sliding-window inference with CSV outputs. Designed for general video action recognition tasks and adaptable to any labeled video dataset.
whwu95 / DSANet【ACMMM'2021】DSANet: Dynamic Segment Aggregation Network for Video-Level Representation Learning
pyf98 / Traffic Accident DetectionIDL course project: Traffic Accident Detection via Deep Learning.
jainsee24 / Parallel Face DetectionImage segmentation is the process of dividing an image into multiple parts. It is typically used to identify objects or other relevant information in digital images. There are many ways to perform image segmentation including Thresholding methods, Color-based segmentation, Transform methods among many others. Alternately edge detection can be used for image segmentation and data extraction in areas such as image processing, computer vision, and machine vision. Image thresholding is a simple, yet effective, way of partitioning an image into a foreground and background. This image analysis technique is a type of image segmentation that isolates objects by converting grayscale images into binary images. Image thresholding is most effective in images with high levels of contrast. Otsu's method, named after Nobuyuki Otsu, is one such implementation of Image Thresholding which involves iterating through all the possible threshold values and calculating a measure of spread for the pixel levels each side of the threshold, i.e. the pixels that either fall in foreground or background. The aim is to find the threshold value where the sum of foreground and background spreads is at its minimum. Edge detection is an image processing technique for finding the boundaries of objects within images. It works by detecting discontinuities in brightness. An image can have horizontal, vertical or diagonal edges. The Sobel operator is used to detect two kinds of edges in an image by making use of a derivative mask, one for the horizontal edges and one for the vertical edges. 1. Introduction Face detection is a computer technology being used in a variety of applications that identifies human faces in digital images. Face detection also refers to the psychological process by which humans locate and attend to faces in a visual scene. Face detection can be regarded as a specific case of object-class detection. In object-class detection, the task is to find the locations and sizes of all objects in an image that belong to a given class. Examples include upper torsos, pedestrians, and cars. Face-detection algorithms focus on the detection of frontal human faces. It is analogous to image detection in which the image of a person is matched bit by bit. Image matches with the image stores in database. Any facial feature changes in the database will invalidate the matching process. 2. Needs/Problems There have been widely applied many researches related to face recognition system. The system is commonly used for video surveillance, human and computer interaction, robot navigation, and etc. Along with the utilization of the system, it leads to the need for a faster system response, such as robot navigation or application for public safety. A number of classification algorithms have been applied to face recognition system, but it still has a problem in terms of computing time. In this system, computing time of the classification or feature extraction is an important thing for further concern. To improve the algorithmic efficiency of face detection, we combine the eigenface method using Haar-like features to detect both of eyes and face, and Robert cross edge detector to locate the human face position. Robert Cross uses the integral image representation and simple rectangular features to eliminate the need of expensive calculation of multi-scale image pyramid. 3. Objectives Some techniques used in this application are 1. Eigen-face technique 2. KLT Algorithm 3. Parallel for loop in openmp 4. OpenCV for face detection. 5. Further uses of the techniques
MengchuanFu / Suspecious Apps DetectionWith the rapid adoption of smartphones, tablets, and mobile apps, they are increasingly becoming part of children’s daily life for amusement and education. According to MarketingProfs, 75% of American children under 8 have access to a smartphone or tablet [1]. According to a survey conducted by PBS KIDS [2] 90% parents with kids between the ages of 2 and 10 state that they believe that educational apps will play an important role in children’s learning in the future. Among smartphone and tablet operating systems, Android and Apple’s iOS dominate the U.S. smartphone market by 53.2 and 41.3 percent, respectively [3]. The growth rate of Apps added to both platform accelerates every year. In 2012, there were a total of 567,322 apps on Google Play platform [4], 723,750 apps on iOS App store [5]. By 2014, the number of apps has doubled, with 1.43 million Apps on Google Play and 1.21 million Apps on iOS [6]. How do parents choose appropriate apps for their kids? PBS’s survey [2] showed that when selecting games and apps for their children, 56% of parents base their decisions on age recommendations. Such decision strategy is far from perfect. There is a rising concern among parents who have experienced unreliable content maturity ratings for mobile apps that result in inappropriate risk exposure for children. In order to help parents determine age-appropriate mobile apps for their children, both Android and iOS apps come with maturity ratings that are similar to the movie and video game industry. Such maturity ratings examine the existence and intensity of mature themes such as mature content, violence, offensive language, sexual content, and drug usage within each app. However, movie and video game industries have official rating organizations such as the Motion Picture Association of America (MPAA) and Entertainment Software Rating Board (ESRB), which set standards for film rating systems – mobile apps do not. Instead of having standard rating rules across platforms, each mobile platform establishes or adopts its own rating policy and rating strategy. iOS’s policy provides four different maturity-rating levels based on the suitable age of audience: “4+,” “9+,” “12+,” and “17+.” In the past, Android rating contained four maturityrating levels: “Everyone,” “Low Maturity,” “Medium Maturity,” and “High Maturity.” Both rating systems classify types of objectionable content into four maturity levels, and their classification rules for each level were similar but still with minor differences. For instance, apps with intense usage of offensive language are rated as “Low Maturity” (maturity level 2) on Android platform, but they are “12+” (maturity level 3) on iOS. Such inconsistency in maturity rating raises much concerns. This inconsistency issue is more challenging now with Android switched to ESRB’s 5 level maturity rating policy with “Everyone,” “Everyone 10+,” “Teen,” “Mature,” and “Adults Only” in May 2015. The inconsistency problem not only lies in policy content, but also is reflected in implementation of maturity rating policy. The main difference between iOS and Android platforms is who determines or reports the actual ratings. iOS rates each app submitted according to its own policies. The maturity ratings for Android apps are purely a result of app developers’ self-report. Developers are required to choose one from the four maturity levels before publishing their apps. After submitting to the Google Play Store, an app is available for download in just a few hours. Google does not verify each app’s maturity rating unless there are a number of user complaints. The public may raise concerns about the authenticity of the maturity ratings of Android apps, but this requires diligent policing on the part of the end user community. In contrast, iOS has a more strict review process for newly released apps. Apple first requires developers to select from a list of objectionable content and indicate the intensity of the content to generate the maturity rating. According to Apple’s “App Store Review Guidelines,” Apple examines the contents of apps and adjusts any inappropriate ratings during a review process before the app becomes available to users [7]. Due to the laxity of Android’s maturity rating policy and the lack of objective judgment of apps’ maturity levels provided by developers, many news articles have recognized the drawbacks of Android’s rating system. They claim that the Android rating policy is unclear, and it is difficult for developers to understand the difference between the four maturity-rating levels [8]. In addition, according to the Washington Post [9] and recent reports from Federal Trade Commission [10, 11], there is a rising concern among parents who have experienced that the maturity ratings of the apps are unreliable. A more critical risk resides in in-app advertisement. Many apps, especially the free ones, are connected to third party advertisements. Neither mobile platforms nor advertising networks apply these maturity policies to restrict the contents of in-app advertisements. As a result, children may still be able to view high maturity contents from in-app advertisements within those apps rated with low maturity. For example, as a 4+ app on iOS platform, Angry Birds should “contain no objectionable materials” as shown in Figure 1. However, it allows a full-screen advertisement (Figure 2) with bloody scenes from a high maturity app appearing inside the app (the source of this bloody scene comes from the 9+ app Blood Brothers containing “Infrequent/Mild Cartoon or Fantasy Violence”). Further, Angry Birds also allows sexual banner advertisements shown on the up-right corner of the screen (Figure 3). As pointed out by the Washington Post, “there have been complaints that violent and sexual ads pop up in some apps aimed at children” [8]. Parents have noticed this issue and are concerned with the objectionable contents of in-app advertisements. However, currently there are no standard rating policy applied to advertisement content and little research investigating this problem. According to our best knowledge, little systematic research has conducted to analyze the problems with apps’ maturity rating policy and their implementation for children’s protection. Thus, the risks associated with content inappropriateness is unknown [12]. This project develops mechanisms to compare, analyze and verify the maturity ratings of mobile apps and in-app advertisements, and investigates the possible reasons behind the inaccurate ratings.
Phirat-Passi / U2Net Image Segmentation ML ModelOne of the most important operations in Computer Vision is Segmentation. Image segmentation is the task of clustering parts of an image together that belong to the same object class. This process is also called pixel-level classification. Here, it involves partitioning images (or video frames) into multiple complex segments or objects in order to produce optimum augmented images utilizing the U2Net and Rembg Image Proceesing Models with OpenVINO toolkit.
Kitware / Burn OutA computer vision toolkit designed to generate pixel-level object classifications, and additionally perform other types of video filtering. Originally created to remove on-screen displays from video.
GuoJiaqi-1020 / MeDiVLADThis is the public repository for the paper "Efficient Lung Ultrasound Severity Scoring Using Dedicated Feature Extractor"
imzhangyd / SAG NetSAG-Net is the official repository of the paper "Using Segment-Level Attention to Guide Breast Ultrasound Video Classification"..
wahidpanda / Introduction To Data Signal And Image Analysis With MATLAB Coursera AnswersIntroduction to Data, Signal, and Image Analysis with MATLAB Welcome to Introduction to Data, Signal, and Image Analysis with MATLAB! MATLAB is an extremely versatile programming language for data, signal, and image analysis tasks, including hundreds if not thousands of functions. With such a comprehensive tool set, knowing where to start can be overwhelming. My goal is to help you learn the basics, with video lessons and assignments that introduce you to the most fundamental functions, show you how to write new code, and demonstrate how to learn how to use functions you have not used before. This course is designed to introduce data, signal, and image processing and analysis to students who have little or no experience with data and signals but have basic programming experience in the MATLAB programming language, for example, those who have completed the Introduction to Programming with Matlab course. The level is targeted at first-year college students and high school seniors, but really this course is suitable for anybody who wants to learn about data and signal analysis and has experience with linear algebra. The length of the course is five weeks. As shown below, that includes four weeks of video lectures plus an extra week for a final project submission. Schedule Week 1: Introduction Week 2: Data analysis in Matlab Week 3: Signal analysis in Matlab Week 4: Image analysis in Matlab Week 5: Course Project Course objectives After completing this course, a learner will be able to use MATLAB to… Understand how signals, images, and data are represented Load and save datasets Visualize high dimensional data Apply machine learning methods for data classification Perform signal frequency analysis Design signal and image filters Process and analyze image content
HemantaIngle / FPGA Based Image RecognitionIn this Project, our main aim is to accelerate the image recognition of CNN (Convolution Neural Network) with the help of a platform deployable on FPGA. CNN focuses on image classification, speech recognition, and video analysis. CNN is accelerated by using GPU (Graphical Processing Unit), which is relatively slow and consumes a high amount of power as CNN requires 20 GFLOPS/image. Also, the CPU acceleration being cheaper as it is readily available on most x86 machines is proportional to power. The modern Application-Specific Chips(ASICS) and the capability of a Field Programmable Gate Array( FPGA ) have power efficiency and faster computation rate over the GPU. With FPGA as a reconfigurable base and parallel architecture, we decided to target the CNN acceleration with an FPGA using Pipe CNN- an algorithm that gets synthesized via HLS (Hardware Level Synthesis Tools) like Intel's Quartus, and Open CL toolkit. Modern Large scale FPGA's like Stratix 10 and Arria 10 have shown a 10 percent less power consumption than GPU's, and it has an added advantage of pipeline parallel architecture and dedicated DSP for faster and efficient computations. The main goal of the Project is to design an OpenCL accelerator that is generic and yet powerful means of improving throughput in inference computations
