search

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
29.35k stars 7.33k forks source link

Modelzoo updates #2869

Open RomRoc opened 3 years ago

RomRoc commented 3 years ago

🚀 Feature

I'm an enthusiastic user of D2, compared with others sw it's more easy, complete, and extensible to use. I have just one question: it seems that modelzoo is not so much updated, compared with other libraries available for ML CV. At the moment the model with best box AP=44.3, is cascade_mask_rcnn_R_50_FPN_3x, even if it's almost 2 years old. Can you please publish some model from recent papers? Now other frameworks offers models to retrain with AP over 50. Thanks

Motivation & Examples

I achieved very good results in object detection, using even TTA and other parameters optimization (NMS_THRESH_TEST, PRE_NMS_TOPK_TEST, ...). I would like to try some other models, from recent papers.

Kevin-Delnoije commented 3 years ago

Totally agree, I like the extensibility and compactness of the library but currently the model zoo is a bit slacking. Personally I would like a close to SOTA segmentation model so you dont have to rely on the deeplab/pointrend projects. An updated object detection model getting closer to SOTA (e.g. FCOS from /projects) would also be very nice.

Perhaps replacing FPN with biFPN and resnet-> [res2net, resnest, effnetv2 ] would already be sufficient for retinanet and sementicsegmentation meta architectures.

Optionally some extra training/ data augmentation techniques could be implemented. see (yolov4/efficientdet)

These models would really help engineers using the library.