-
Download CityPersons coco2014 CrowdHuman CUHKocclusion ETHZ VOCdevkit dataset to train.
-
convert annotations as follow: retina-train.txt content:
892 445 913 498 #lefttop point right bottom point 901 443 935 498 1844 436 1888 542
1290 425 1315 486
$ train.py [-h] [data_path DATA_PATH] [--batch BATCH] [--epochs EPOCHS] [--shuffle SHUFFLE] [img_size IMG_SIZE] [--verbose VERBOSE] [--save_step SAVE_STEP] [--eval_step EVAL_STEP] [--save_path SAVE_PATH] [--depth DEPTH]
For multi-gpus training, run:
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train.py
missing due to other reason
the improvement of AP is little compare to face detect. the reason maybe:
- the sample has more mistake in label. so focal loss will learn more bad thing.
- pedestrian box has more background and shield. the detection is more difficult.
- [ ]
refrence: https://github.com/supernotman/RetinaFace_Pytorch.git