- …
- …
#245
summarized by : Anonymous
どんな論文か？
A good solution for how to spatially sample features and how to process sampled features for DETR-like detectors; no need encoder, no need FPN;
新規性
adaptive 3D feature sampling (positional vector + learned x, y, z offset from content vector); adaptive mixing (channel + spatial mixing with weights learned from content vector), no self-attention
結果
superior performance, especially on small object detection, with less computational cost compared to other query-based detectors
その他（なぜ通ったか？等）
- …
- …