Small Object Detection Based on SSD Pipeline

Introduction

This is our final project of EI339 on small object detection. Our implementation references ssd.pytorch and ssds.pytorch. Several modifications on SSD are implemented:

Add more anchor boxes
Augment small objects with cut-and-paste
Add deconvolution layers (not work now)
Add non-local block
Feature-Fused SSD
Receptive Field Block

Usage

To train ssd with added anchor boxes and non-local block:

python ssd/train.py

To train fssd or rfb:

python ssds.pytorch/train.py --cfg=ssds.pytorch/experiments/cfgs/fssd_vgg16_train_voc.yml

python ssds.pytorch/train.py --cfg=ssds.pytorch/experiments/cfgs/rfb_resnet50_train_voc.yml

To evaluate ssd with added anchor boxes and non-local block:

python ssd/eval.py

Evaluation on small objects:

python ssd/eval_small.py

To visualize bounding box on VOC2007 test, ssd/JPEGImages/ contains VOC2007 test images:

python ssd/viz_bb.py

Performance

Here is performance comparison in our experiment.

| Method | mAP | mAP(small) | mREC | mREC(small)| FPS | | :-----| :---- | :---- | :-----| :---- | :---- | | SSD | 77.3 | 52.6 | 93.4 | 83.4 | 23 | | SSD-ab | 77.7 | 53.2 | 92.9 | 82.3 | 24 | | SSD-aug | 77.1 | 52.5 | 93.1 | 82.2 | 14 | | SSD-nlb | 78.0 | 52.9 | 93.8 | 82.8 | 24 | | FSSD | 77.6 | 56.8 | 92.1 | 82.5 | 21 | | RFB | 80.9 | 59.8 | 95.6 | 88.2 | 7 |

where SSD-ab, SSD-aug, SSD-nlb denote SSD with added anchor boxes, SSD with small object augmentatio, SSD with non-local blocks respectively. All trained models are available on https://drive.google.com/drive/folders/15Dm4MefcpbRkDMVT1ZXtmL5Ea6v4fUTZ?usp=sharing.