YOLOv3 from scratch

This is the mid-term project for DATA130051 Computer Vision. 🐮🐮🐮

Full report <here>.

YOLOv3 🚀 is an open-source object detection architecture first proposed by Joseph Redmon[1]. I managed to reproduce the network from scratch with the help of Aladdin Persson[2] and his excellent video tutorials[3]. 🤳 Basically, I reimplemented YOLOv3 model as well as the utils and metrics for detections. Also, a detector class was written to provide a shortcut for video detections. Again, thanks a lot for those who contribute to this open-source society! 🤗

Quick Start 🌟

Install 😎

Python>=3.6.0 is required with all requirements.txt installed including PyTorch>=1.9.0:

git clone https://github.com/super-dainiu/yolov3
cd yolov3
pip install -r requirements.txt

Download weights 🤪

Please follow the guide in ./pretrained directory.

Download dataset 🥵

Please follow the guide in ./PASCAL_VOC directory. Again credit to Aladdin Persson[2] for his excellent tutorial. 👍

Inference 👾

After downloading the weights, you can do detection for images, gifs, or mp4 using main.py.

usage: main.py [-h] --weights WEIGHTS [--save SAVE] [--save_dir SAVE_DIR] [--conf_thres CONF_THRES] [--iou_thres IOU_THRES] [--max_det MAX_DET] [--target TARGET [TARGET ...]] [--camera CAMERA] [--samples SAMPLES [SAMPLES ...]]

Use YOLOv3

optional arguments:

-h, --help show this help message and exit

--weights WEIGHTS Weight directory

--save SAVE Save results

--save_dir SAVE_DIR Save directory

--conf_thres CONF_THRES Confidence threshold

--iou_thres IOU_THRES IOU threshold

--max_det MAX_DET Maximum detection per frame

--target TARGET [TARGET ...] Targets (i.e. person), * for all classes

--camera CAMERA Use your camera

--samples SAMPLES [SAMPLES ...] Sample images (ends with .jpg, .png, .gif, .mp4)

You might do a simple detection with the following bash command.

python main.py --weights pretrained/yolov3.pth.tar --sample samples/video.mp4 samples/image_1.jpg samples/image_2.png samples/gif_1.gif --save True --save-dir outputs --target person car --conf_thres 0.7 --iou_thres 0.3 --max_det 10

Or activate your camera for detection.

python main.py --weights pretrained/yolov3.pth.tar --camera True

Results 🙀

PASCAL VOC metrics (mAP at .5 IOU) 🌟

Class	mAP
aeroplane	58.018
bicycle	60.933
bird	34.835
boat	32.412
bottle	17.423
bus	63.038
car	69.853
cat	57.547
chair	29.547
cow	44.667
diningtable	48.446
dog	48.934
horse	59.299
motorbike	63.281
person	52.751
pottedplant	19.121
sheep	48.951
sofa	51.492
train	56.918
tvmonitor	55.288
TOTAL mAP	48.638

Example Outputs 💩

Transformer

CR7 Celebration

Maguire and Lingod

The dog

The cat

Car