DAFormer icon indicating copy to clipboard operation
DAFormer copied to clipboard

imnet_feature_dist_classes=[6, 7, 11, 12, 13, 14, 15, 16, 17, 18]

Open creater-zq opened this issue 3 years ago • 1 comments

Hello!

  • How to choose these classes in experiment? such astraffic lighttraffic sign, etc.
  • Because in training, I observed src.loss_imnet_feat_dist: nan.
  • thanks!

creater-zq avatar Jul 25 '22 03:07 creater-zq

Hello,

  1. These classes are chosen to be thing-classes (countable objects such as car or traffic light) in contrast to stuff-classes (background regions such as road). Please, also have a look at Sec. B in the Supplement: "For the distinction of thing- and stuff-classes, we follow the definition by Caesar et al. [4]. Applied to Cityscapes, thing-classes are traffic light, traffic sign, person, rider, car, truck, bus, train, motorcycle, and bicycle and stuff-classes are road, sidewalk, building, wall, fence, pole, vegetation, terrain, and sky." [4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Coco-stuff: Thing and stuff classes in context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1209–1218, 2018.
  2. Please, have a look at Issue #11.

lhoyer avatar Jul 25 '22 07:07 lhoyer