Learning Spatial Fusion for Single-Shot Object Detection

By Songtao Liu, Di Huang, Yunhong Wang

Introduction

In this work, we propose a novel and data driven strategy for pyramidal feature fusion, referred to as adaptively spatial feature fusion (ASFF). It learns the way to spatially filter conflictive information to suppress the inconsistency, thus improving the scale-invariance of features, and introduces nearly free inference overhead. For more details, please refer to our arXiv paper.

Updates:

• YOLOX is here!, come and use the stronger YOLO!

• Add MobileNet V2!

  • The previous models actually are all trained with the wrong anchor setting, we fix the error on mobileNet model.
  • We currently not support rfb, dropblock and Feature Adaption for mobileNet V2.
  • FP16 training for mobileNet is not working now. I didn't figure it out.
  • FP16 testing for mobileNet drops about 0.2 mAP.

• Add a demo.py file

• Faster NMS (adopt official implementation)

COCO

Citing

Please cite our paper in your publications if it helps your research:

@article{liu2019asff,
    title = {Learning Spatial Fusion for Single-Shot Object Detection},
    author = {Songtao Liu, Di Huang and Yunhong Wang},
    booktitle = {arxiv preprint arXiv:1911.09516},
    year = {2019}
}

Contents

1. Installation
2. Datasets
3. Training
4. Evaluation
5. Models

Installation

• Install PyTorch-1.3.1 by selecting your environment on the website and running the appropriate command.
• Clone this repository.
  • Note: We currently only support PyTorch-1.0.0+ and Python 3+.
• Compile the DCN layer (ported from DCNv2 implementation):

./make.sh

Prerequisites

• We also use apex, numpy, opencv, tqdm, pyyaml, matplotlib, scikit-image...

  • Note: We use apex for distributed training and synchronized batch normalization. For FP16 training, since the current apex version have some issues, we use the old version of FP16_Optimizer, and split the code in ./utils/fp_utils.

• We also support tensorboard if you have installed it.

Demo

python demo.py -i /path/to/your/image \
--cfg config/yolov3_baseline.cfg -d COCO \
--checkpoint /path/to/you/weights --half --asff --rfb -s 608

• Note:
  • -i, --img: image path.
  • --cfg: config files.
  • -d: choose datasets, COCO or VOC.
  • -c, --checkpoint: pretrained weights.
  • --half: FP16 testing.
  • -s: evaluation image size, from 320 to 608 as in YOLOv3.

Datasets

Note: We currently only support COCO and VOC.
To make things easy, we provide simple COCO and VOC dataset loader that inherits torch.utils.data.Dataset making it fully compatible with the torchvision.datasets API.

Moreover, we also implement the Mix-up strategy in BoFs and distributed random resizing in YOLov3.

COCO Dataset

Install the MS COCO dataset at /path/to/coco from official website, default is ./data/COCO, and a soft-link is recommended.

ln -s /path/to/coco ./data/COCO

It should have this basic structure

$COCO/
$COCO/annotations/
$COCO/images/
$COCO/images/test2017/
$COCO/images/train2017/
$COCO/images/val2017/

The current COCO dataset has released new train2017 and val2017 sets, and we defaultly train our model on train2017 and evaluate on val2017.

VOC Dataset

Install the VOC dataset as ./data/VOC. We also recommend a soft-link:

ln -s /path/to/VOCdevkit ./data/VOC

Training

• First download the mix-up pretrained Darknet-53 PyTorch base network weights at: https://drive.google.com/open?id=1phqyYhV1K9KZLQZH1kENTAPprLBmymfP
  or from our BaiduYun Driver

• For MobileNetV2, we use the pytorch official weights (change the key name to fit our code), or from our BaiduYun Driver

• By default, we assume you have downloaded the file in the ASFF/weights dir:

• Since random resizing consumes much more GPU memory, we implement FP16 training with an old version of apex.

• We currently ONLY test the code with distributed training on multiple GPUs (10 2080ti or 4 Tesla V100).

• To train YOLOv3 baseline (ours) using the train script simply specify the parameters listed in main.py as a flag or manually change them on config/yolov3_baseline.cfg:

python -m torch.distributed.launch --nproc_per_node=10 --master_port=${RANDOM+10000} main.py \
--cfg config/yolov3_baseline.cfg -d COCO --tfboard --distributed --ngpu 10 \
--checkpoint weights/darknet53_feature_mx.pth --start_epoch 0 --half --log_dir log/COCO -s 608 

• Note:

  • --cfg: config files.

  • --tfboard: use tensorboard.

  • --distributed: distributed training (we only test the code with distributed training)

  • -d: choose datasets, COCO or VOC.

  • --ngpu: number of GPUs.

  • -c, --checkpoint: pretrained weights or resume weights. You can pick-up training from a checkpoint by specifying the path as one of the training parameters (again, see main.py for options)

  • --start_epoch: used for resume training.

  • --half: FP16 training.

  • --log_dir: log dir for tensorboard.

  • -s: evaluation image size, from 320 to 608 as in YOLOv3.

• To train YOLOv3 with ASFF or ASFF*, you only need add some addional flags:

python -m torch.distributed.launch --nproc_per_node=10 --master_port=${RANDOM+10000} main.py \
--cfg config/yolov3_baseline.cfg -d COCO --tfboard --distributed --ngpu 10 \
--checkpoint weights/darknet53_feature_mx.pth --start_epoch 0 --half --asff --rfb --dropblock \
--log_dir log/COCO_ASFF -s 608 

• Note:
  • --asff: add ASFF module on YOLOv3.
  • --rfb: use RFB moduel on ASFF.
  • --dropblock: use DropBlock.

Evaluation

To evaluate a trained network, you can use the following command:

python -m torch.distributed.launch --nproc_per_node=10 --master_port=${RANDOM+10000} eval.py \
--cfg config/yolov3_baseline.cfg -d COCO --distributed --ngpu 10 \
--checkpoint /path/to/you/weights --half --asff --rfb -s 608

• Note:
  • --vis: Visualization of ASFF.
  • --testset: evaluate on COCO test-dev.
  • -s: evaluation image size.

By default, it will directly output the mAP results on COCO val2017 or VOC test 2007.

Models

• yolov3 mobilenetv2 (ours)weights baiduYun training tfboard log

• yolov3 mobilenetv2 +asff weights baiduYun training tfboard log

• yolov3_baseline (ours) weights baiduYun training tfboard log

• yolov3_asff weights baiduYun training tfboard log

• yolov3_asff* (320-608) weights baiduYun

• yolov3_asff* (480-800) weights baiduYun