facebookresearch / detectron2

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.
https://detectron2.readthedocs.io/en/latest/
Apache License 2.0
30.24k stars 7.44k forks source link

Pretrained models for Keypoint + Segmentation tasks #3099

Open saidaltindis opened 3 years ago

saidaltindis commented 3 years ago

Pretrained models for Keypoint + Segmentation tasks

Although there are multiple pretrained models provided for segmentation and keypoint tasks separately, there is no pretrained model that trained for multi-task.

Providing such models would be an important baseline for the people that re-trained their model since it is possible to train such models using detectron2 (#17).

VGrondin commented 3 years ago

The keypoints model are multi-task, you just need to turn it on e.g.:

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x.yaml"))
cfg.MODEL.MASK_ON = True
cfg.MODEL.KEYPOINT_ON = True 
saidaltindis commented 3 years ago

Are you sure? They did not report the mask score for keypoint models and in metrics, mask loss is not present.

https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_50_FPN_1x/137261548/metrics.json

chenzhutian commented 2 years ago

After setting cfg.MODEL.MASK_ON = True, the output of the mask is just noise. I guess the pre-trained models dont support keypoint detection and segmentation in the same time.

soulslicer commented 2 years ago

Did you every figure this out?