Closed subinium closed 3 years ago
다른 task에서만 적용되던 End-to-End를 Object Detection에서 적용
object detection에서 어려운 것은 predicted objects와 ground truth와의 매칭
Hungrarian Loss에서 no-object 항은 cross entropy에서 1/10로 계산
Bounding Box Loss : l1 loss + genrealized IoU loss
positional encoding + cnn-based fetures로 sequence 제공 (Encoder)
Decoder에서 나온 결과 + FFN으로 결과 예측
실험 결과
https://arxiv.org/abs/2005.12872
DETR
읽기에 앞서 기존 object detection 방법을 먼저 알고 보는 게 좋을 것 같아서 (오디알못...)
Jusin Johnson : Lecture 15: Object Detection