Keras implementation of of paper: A Self-Training Approach for Point-Supervised Object Detection and Counting in Crowds. The code is developed based on CSP. This work is accepted by IEEE TIP 2021.
- Training and Testing Code of WiderFace
- Training and Testing Code of ShanghaiTech
In this article, we propose a novel self-training approach named Crowd-SDNet that enables a typical object detector trained only with point-level annotations (i.e., objects are labeled with points) to estimate both the center points and sizes of crowded objects.
This is achieved by the proposed locally-uniform distribution assumption, the crowdedness-aware loss, the confidence and order-aware refinement scheme, and the effective decoding method, which promote the detector to generate accurate bounding boxes in a coarse-to-fine and end-to-end manner.
-
Python 2.7
-
Tensorflow 1.14.1
-
Keras 2.0.6
The libraries are shown in
./requirements.txt.
- Clone this repo in the directory.
git clone https://github.com/WangyiNTU/Point-supervised-crowd-detection.git
- Install the requirments.
cd Point-supervised-crowd-detection
pip install -r requirements.txt
-
Dataset downloading.
1.1 For the WiderFace dataset, we download the training and validation images and annotations at WiderFace and put it in
./data/WIDERwith the folder tree:1.2 For the ShanghaiTech dataset, we download the training and validation images and annotations at ShanghaiTech and put it in
./data/part_Aor./data/part_Bwith the folder tree:-- WIDER |-- WIDER_train |-- images |-- 0--Parade |-- 1--Handshaking |-- ... |--WIDER_val |-- images |-- 0--Parade |-- 1--Handshaking |-- ... |--wider_face_split |-- wider_face_train_bbx_gt.txt |-- wider_face_val_bbx_gt.txt |-- ... --part_A or part_B |-- train_data |-- images |-- IMG_1.jpg |-- ... |-- ground-truth |-- GT_IMG_1.mat |-- ... |-- val_data |-- images |-- ground-truth |-- test_data |-- images |-- ground-truth -
Dataset preparation.
Prepare an annotation-cache file by the locally-uniform distribution assumption (LUDA) method, which generates the pseudo object sizes. (Change WiderFace/ShanghaiTech dataset folder and split (train or val) in the following code.)
By default, we assume the cache files is stored in
./data/cache/.python LUDA_generate_wider.pyor (Change Line 13-15 to generate train, train_val, and val cache, respectively),
python LUDA_generate_SH.py -
Initialized models preparation.
We use the backbone ResNet-50 in our experiments. By default, we assume the weight files is stored in
./data/models/.
-
Train on WiderFace.
Follow the ./train_wider.py to start training. The weight files of all epochs will be saved in
./output/valmodels/wider.Optionally, you should set the training parameters in ./keras_csp/config.py or replace some parameters in train_wider.py.
-
Train on ShanghaiTech.
Follow the ./train_ShanghaiTechA.py or ./train_ShanghaiTechB.py to start training. The weight files of all epochs will be saved in
./output/valmodels/.
Since the codes are slightly different from the paper version with fine-tuned parameters and random seeds, the results may be better or worse. Re-implementation of paper's results can refer to the trained model Models.
-
Test on WiderFace.
Follow the ./test_wider.py to get the detection results. The results will be saved in
./output/valresults/wider. Without Training, the test code can be performed with pretrained model from Models.We adopt the similar multi-scale testing as CSP and Point in, Box Out. Since using multi-scale testing and there is no optimization of soft nms (cpu python version), it will take several hours to run the val set. Some scales like [0.75, 1.25, 1.75] can be removed to increase inference speed with a little performance decrease.
-
Test on ShanghaiTech. Follow the ./test_ShanghaiTechA.py or ./test_ShanghaiTechB.py to get the detection results. The results will be saved in
./output/valresults/. Without Training, the test code can be performed with pretrained model from Models.
-
Calculate AP scores for WiderFace dataset.
Download Evaluation tool for the official website of WiderFace.
Follow the wider_eval.m to obtain AP scores and charts on the easy, medium, and hard subsets. (Specify your prediction directory: pred_dir)
-
Counting by detection.
Get the MAE, MSE, NAE on the WiderFace val set / ShanghaiTech test set by running
python tools/counting_by_detection.pySpecify 'part' to the dataset name (wider, part_A, part_B) in Line 101.
-
Plot predicted detections.
Plot detections on the image files by running
python tools/plot_wider_detections.pyChange thr and epoch of Line 13-14 for your trained model. or,
python tools/plot_ShanghaiTech_detections.pyChange part, thr and epoch of Line 12-23 for your trained model.
Qualitative results on (a) WiderFace, (b) SHA, (c) SHB, (d) PUCPR+, and (e) CARPK. The top row shows the ground-truth boxes or points, and counts. The bottom row shows the bounding boxes and counts predicted by our approach.
Comparisons between the bounding boxes (at top row) produced by our approach and the density maps (at bottom row) produced by CSRNet.
To reproduce the results in our paper, we have provided the models. You can download them through Onedrive.
-
For WiderFace
Download wider.zip, put it in the
./output/valmodels, and unzip to./output/valmodels/wider/h/off/ -
For ShanghaiTech
Download part_A model ShanghaiTechA.zip, put it in the
./output/valmodels, and unzip to./output/valmodels/ShanghaiTechA/h/nooff/Download part_B model ShanghaiTechB.zip, put it in the
./output/valmodels, and unzip to./output/valmodels/ShanghaiTechB/h/nooff/
If you find this research project is useful for your research, please star ⭐ this repository and consider citing 📝:
@article{wang2021self_training,
title={A Self-Training Approach for Point-Supervised Object Detection and Counting in Crowds},
author={Wang, Yi and Hou, Junhui and Hou, Xinyu and Chau, Lap-Pui},
journal={IEEE Transactions on Image Processing},
volume={30},
pages={2876--2887},
year={2021},
publisher={IEEE}
}