Why 3 classes instead of 2?

@liuruijin17 Hi! I have a small question: would you kindly tell me why use 3 classes? Labels are all 1(lane)/2(no object), it seems 2 classes should suffice. Then there is another question, maybe one can directly use BCE loss here with only 1-dim output with class_embed? I’ve checked the DETR repo for discussions on single class detection and still can’t figure out why you use 3-dim output at class_embed.

EDIT: Is it related to COCO index starts at 1?

[liuruijin17]

https://github.com/liuruijin17/LSTR/blob/6044f7b2c5892dba7201c273ee632b4962350223/models/py_utils/detr_loss.py#L46 She changes the second argument to be zero since the above line would initialize target_classes as a full ‘2’ tensor (‘2’ for background)

https://github.com/liuruijin17/LSTR/blob/6044f7b2c5892dba7201c273ee632b4962350223/models/py_utils/detr_loss.py#L33 then changes the empty_weight is initialized by torch.ones(self.num_classes)

The difference is (1) 3 classes output, ‘1’: lane, ‘2’ background, ‘0’ no supervision (2) 2 classes output, ‘1’: lane, ‘0’ background

She trained on another dataset, so I would train this version when our GPUs have free spaces. If the accuracy also drops, then there still may have bugs that should be fixed.

[voldemortX]

@liuruijin17 Thanks! So 2/3 classes should not affect best hyper-parameters. Maybe my slightly lower performance comes from not freezing the first BN layer.