🌟💡 YOLOv5 Study: Hybrid Autolabelling

[glenn-jocher]

Study 🤔

I had an idea to hybrid label the COCO train set, and then retrain on it. This discussion is mainly a placeholder for later results on this.

In the process of this I realized there is no easy way to run test.py on the train set, so I created PR #2453 to allow for selection of any of the 3 sets defined in a data.yaml file (train, val, test) by passing them to the --task argument. After this PR hybrid autolabeling command is:

python test.py --task train --data coco.yaml --weights yolov5x6.pt --img 1280 --iou 0.7 --conf 0.90 --save-txt --save-hybrid --project hybrid --name conf_90

Results 😃

TODO

[robotaiguy]

I'm planning on doing similar, because a while back I trained a model with images from both COCO2017 and Open images, by using Fiftyone to download ONLY images with Person labels, and I noticed that there exists plenty of opportunity for helping those datasets, especially Open Images. So I'll be using the biggest, baddest, Yolov5x6 model an inferencing them all with --save-txt enabled.
I should be able to knock it out pretty quick with a 3090 and an A100. I was actually using Yolov5 for hybrid autolabeling this morning.

[glenn-jocher]

@robotwhispering yes hybrid autolabelling is the best way to improve an existing dataset without touching the existing labels.

[PraveenMNaik]

@robotwhispering Sir, pls show us how to configure for autolabelling

[glenn-jocher]

@PraveenMNaik you can autolabel image with detect.py or val.py:

python detect.py --save-txt 
python val.py --data DATA.yaml --save-txt

[spacewalk01]

May refer to this repo for autolabeling: https://github.com/cnyvfang/labelGo-Yolov5AutoLabelImg