CenterM2: Pursuing Better Results via Macro and Micro Regulation

Abstract

Object detectors gradually tend to imitate the human visual system since the way of understanding the scene from local information to global percepts made the human visual system both efficient and accurate. However, most existing detectors attend to solely focus on obvious regions of the scene which are not always regions of interest. In this work, a new anchor-free object detector, CenterM2, is proposed to focus on the regions of interest via macro and micro regulation. At the macro level, a multi-scale fusion module is utilized to enrich the semantic and edge position information for the detector. At the micro level, semantic features of different receptive fields in the network are adaptively fused by using switchable atrous convolution, and an adaptive feature fusion module is exploited to focus on the regions of interest between low-level and high-level feature maps. On the PASCAL VOC2007 dataset, CenterM2 achieves 73.3% mAP (mean average precision) at 71 FPS (frames per second) and 85.5% mAP at 27 FPS. Furthermore, it can get a better trade-off between detection speed and accuracy, such as 26.9% AP (average precision) at 71 FPS, 42.0% AP at 24 FPS, and 45.3% AP with multi-scale testing at 3 FPS on the MS COCO dataset.

Main results

Object Detection on COCO validation

Backbone	AP / FPS	Multi-scale AP / FPS	Download
ResNet-18	33.6 / 26	-	res18_sac_coco
DLA-34	39.4 / 25	-	dla34_sac_coco
ResNet-101	40.5 / 11	44.8 / 2	res101_sac_coco
CSPDarknet-53	42.6 / 24	45.6 / 4	csp53_coco_val

Object Detection on PASCAL VOC2007

Backbone	mAP	FPS	Multi-scale mAP / FPS	Download
Darknet-Tiny	73.3	71	-	darktiny_voc_73.3
ResNet-18	78.4	32	-	res18_sac_voc_78.4
DLA-34	81.7	27	-	dla34_sac_voc_81.7
Darknet-53	81.6	31	83.4 / 4	dark53_voc_81.6
ResNet-101	82.0	12	83.8 / 2	res101_sac_voc_82.0
CSPDarknet-53	85.5	27	86.0 / 4	csp53_voc_85.5

Installation

[Optional but recommended] create a new conda environment.
```
conda create --name CenterM2 python=3.7
```
And activate the environment.
```
conda activate CenterM2
```

Install pytorch1.3.0:

conda install pytorch=1.3.0 torchvision -c pytorch

Install COCOAPI:

# COCOAPI=/path/to/clone/cocoapi
git clone https://github.com/cocodataset/cocoapi.git $COCOAPI
cd $COCOAPI/PythonAPI
make
python setup.py install --user

Clone this repo:

CenterM2_ROOT=/path/to/clone/CenterM2
git clone https://github.com/michunf/CenterM2 $CenterM2_ROOT

Install the requirements
```
pip install -r requirements.txt
```

Compile deformable convolutional.

cd $CenterM2_ROOT/src/lib/models/networks/DCNv2
./make.sh

Compile NMS.

cd $CenterM2_ROOT/src/lib/external
make

Download pertained models for detection and move them to $CenterM2_ROOT/models/.

Use CenterM2 (refer to CenterNet for more detailed instructions)

We support demo for image/ image folder, video, and webcam.

First, download the models (By default, csp53_coco_val and cspdarknet53--Tianxiaomo for detection and put them in CenterM2_ROOT/models/.

For object detection on images/ video, run:

python demo.py ctdet --demo /path/to/image/or/folder/or/video --load_model ../csp53_coco_val.pth

For webcam demo, run

python demo.py ctdet --demo webcam --load_model ../models/csp53_coco_val.pth

You can add --flip_test and --nms for flip test. You can add --debug 2 to visualize the heatmap outputs.

Dataset preparation

If you want to reproduce the results in the paper for benchmark evaluation and training, you will need to setup dataset.

COCO

Download the images (2017 Train, 2017 Val, 2017 Test) from coco website.
Download annotation files (2017 train/val and test image info) from coco website.

Place the data (or create symlinks) to make the data folder like:

${CenterM2_ROOT}
|-- data
`-- |-- coco
    `-- |-- annotations
        |   |-- instances_train2017.json
        |   |-- instances_val2017.json
        |   |-- person_keypoints_train2017.json
        |   |-- person_keypoints_val2017.json
        |   |-- image_info_test-dev2017.json
        |---|-- train2017
        |---|-- val2017
        `---|-- test2017

Pascal VOC

Run

cd $CenterM2_ROOT/tools/
bash get_pascal_voc.sh

The above script includes:
- Download, unzip, and move Pascal VOC images from the VOC website.
- Download Pascal VOC annotation in COCO format (from Detectron).
- Combine train/val 2007/2012 annotation files into a single json.

Move the created voc folder to data (or create symlinks) to make the data folder like:

${CenterM2_ROOT}
|-- data
`-- |-- voc
    `-- |-- annotations
        |   |-- pascal_trainval0712.json
        |   |-- pascal_test2017.json
        |-- images
        |   |-- 000001.jpg
        |   ......
        `-- VOCdevkit

The VOCdevkit folder is needed to run the evaluation script from faster rcnn.

Benchmark Evaluation and Training

First, download the pretrained weights for darknet:

darktiny, darknet53--Kissrabbit, and cspdarknet53--Tianxiaomo

For training on a single GPU, run

python main.py ctdet --exp_id your_file_name --dataset coco --arch your_model_name --batch_size 32 --num_workers 4 --lr 1.25e-4 --lr_step 90,120 --num_epochs 140 --save_all

For training on multiple GPUs, run

python main.py ctdet --exp_id your_file_name --dataset coco --arch your_model_name --gpus 0,1,2,3,... --batch_size 32*num_GPUs --num_workers 4*num_GPUs --lr 1.25e-4*num_GPUs --lr_step 90,120 --num_epochs 140 --save_all

To evaluate COCO object detection (all models are trained on COCO train 2017 and evaluated on val 2017), run

python test.py ctdet --exp_id your_file_name --flip_test --nms --load_model ../models/csp53_coco_val.pth

To evaluate Pascal object detection (all models are trained on trainval 07+12 and tested on test 2007), for 384x384, run

python test.py ctdet --exp_id your_file_name --flip_test --keep_res --nms --load_model ../models/darktiny_voc_73.3.pth

for 512x512, run

python test.py ctdet --exp_id your_file_name --flip_test --nms --load_model ../models/model_voc_130.pth

for mutile scale test, add --test_scales 0.75,1,1.25,1.5,1.75,2 and --keep_res.

ChunfengMI / CenterM2