CSL-YOLO: A New Lightweight Object Detection System for Edge Computing
This project provides a SOTA level lightweight YOLO called "Cross-Stage Lightweight YOLO"(CSL-YOLO),
it is achieving better detection performance with only 43% FLOPs and 52% parameters than Tiny-YOLOv4.
Paper Link: https://arxiv.org/abs/2107.04829
Requirements
How to Get Started?
#Predict
python3 main.py -p cfg/predict_coco.cfg
#Train
python3 main.py -t cfg/train_coco.cfg
#Eval
python3 main.py -ce cfg/eval_coco.cfg
WebCam DEMO(on CPU)
This DEMO runs on a pure CPU environment, the CPU is I7-6600U(2.6Ghz~3.4Ghz), the model scale is 224x224, and the FPS is about 10.
Please execute the following script to get this DEMO, the "camera_idx" in the cfg file represents the camera number you specified.
#Camera DEMO
python3 main.py -d cfg/demo_coco.cfg
More Info
Change Model Scale
The model's default scale is 224x224, if you want to change the scale to 320~512,
please go to cfg/XXXX.cfg and change the following two parts:
# input_shape=[512,512,3]
# out_hw_list=[[64,64],[48,48],[32,32],[24,24],[16,16]]
# input_shape=[416,416,3]
# out_hw_list=[[52,52],[39,39],[26,26],[20,20],[13,13]]
# input_shape=[320,320,3]
# out_hw_list=[[40,40],[30,30],[20,20],[15,15],[10,10]]
input_shape=[224,224,3]
out_hw_list=[[28,28],[21,21],[14,14],[10,10],[7,7]]
weight_path=weights/224_nolog.hdf5
|
| 224 to 320
V
# input_shape=[512,512,3]
# out_hw_list=[[64,64],[48,48],[32,32],[24,24],[16,16]]
# input_shape=[416,416,3]
# out_hw_list=[[52,52],[39,39],[26,26],[20,20],[13,13]]
input_shape=[320,320,3]
out_hw_list=[[40,40],[30,30],[20,20],[15,15],[10,10]]
# input_shape=[224,224,3]
# out_hw_list=[[28,28],[21,21],[14,14],[10,10],[7,7]]
weight_path=weights/320_nolog.hdf5
Fully Dataset
The entire MS-COCO data set is too large, here only a few pictures are stored for DEMO,
if you need complete data, please download on this page.
Our Data Format
We did not use the official format of MS-COCO, we expressed a bounding box as following:
[ left_top_x<float>, left_top_y<float>, w<float>, h<float>, confidence<float>, class<str> ]
The bounding boxes contained in a picture are represented by single json file.
For detailed format, please refer to the json file in "data/coco/train/json".
AP Performance on MS-COCO
For detailed COCO report, please refer to "mscoco_result".
TODOs
- Improve the calculator script of FLOPs.
- Using Focal Loss will cause overfitting, we need to explore the reasons.
5 Oct 18, 2022
3 Jan 26, 2022
39 Dec 28, 2022
865 Dec 24, 2022
2 Mar 16, 2022
567 Dec 26, 2022
7 Aug 16, 2022
1 Nov 12, 2021
6 Dec 8, 2022
342 Jan 3, 2023
2 Jul 6, 2022
60 May 29, 2022
16 Oct 14, 2022
349 Dec 8, 2022
88 Dec 30, 2022
152 Jan 2, 2023
405 Jan 4, 2023
2.8k Jan 8, 2023