COCO model does not recognize pedestrians very well

WisconsinAIVision / yolact_edge

The first competitive instance segmentation approach that runs on small edge devices at real-time speeds.

MIT License

1.27k stars 273 forks source link

COCO model does not recognize pedestrians very well #149

Open hoanglongmoovita opened 3 years ago

hoanglongmoovita commented 3 years ago

Hi,

I am trying out the yolact edge model on a car. It works great in detecting vehicles, but it hardly ever detects pedestrians, even at close range. Is this because your dataset doesn't include many pedestrians?

I have tried the 2 R-101-FPN models for COCO and for YoutubeVIS and result is the same

Environment:

OS: Ubuntu 20.04
GPU: RTX 3060 Ti
CUDA Version 11.3

KastanDay commented 2 years ago

I'm experiencing the same issue. Only one of the pretrained models detected people at all, and even then it was quite poor. The rest of them often failed to return any predictions for people at all.

I had the best luck with yolact_edge_54_800000.pth but your mileage may vary. Detectron2 from Facebook works great for people, but I can't get it to work on Jetson Xaiver.

haotian-liu commented 2 years ago

I have not specifically conducted experiment on the person class. The performance of YolactEdge (without video) is highly dependent on that of YOLACT, though. Besides, I have seen people working on person datasets as well #47, seems that it detects person quite well when it is specially trained on the person class -- which might suggests either a re-calibrated dataset or some data augmentation might help with the person detection case.