File Download
There are no files associated with this item.
Supplementary
-
Citations:
- Appears in Collections:
Conference Paper: What Makes for End-to-End Object Detection?
Title | What Makes for End-to-End Object Detection? |
---|---|
Authors | |
Issue Date | 2021 |
Publisher | ML Research Press. The Journal's web site is located at http://proceedings.mlr.press/ |
Citation | The 38th International Conference on Machine Learning (ICML), Virtual Conference, 18-24 July 2021. In Proceedings of Machine Learning Research (PMLR), v. 139: Proceedings of ICML 2021, p. 9934-9944 How to Cite? |
Abstract | Object detection has recently achieved a breakthrough for removing the last one non-differentiable component in the pipeline, Non-Maximum Suppression (NMS), and building up an end-to-end system. However, what makes for its one-to-one prediction has not been well understood. In this paper, we first point out that one-to-one positive sample assignment is the key factor, while, one-to-many assignment in previous detectors causes redundant predictions in inference. Second, we surprisingly find that even training with one-to-one assignment, previous detectors still produce redundant predictions. We identify that classification cost in matching cost is the main ingredient: (1) previous detectors only consider location cost, (2) by additionally introducing classification cost, previous detectors immediately produce one-to-one prediction during inference. We introduce the concept of score gap to explore the effect of matching cost. Classification cost enlarges the score gap by choosing positive samples as those of highest score in the training iteration and reducing noisy positive samples brought by only location cost. Finally, we demonstrate the advantages of end-to-end object detection on crowded scenes. |
Description | Poster Session 2 |
Persistent Identifier | http://hdl.handle.net/10722/301312 |
ISSN |
DC Field | Value | Language |
---|---|---|
dc.contributor.author | Sun, P | - |
dc.contributor.author | Jiang, Y | - |
dc.contributor.author | Xie, E | - |
dc.contributor.author | Shao, W | - |
dc.contributor.author | Yuan, Z | - |
dc.contributor.author | Wang, C | - |
dc.contributor.author | Luo, P | - |
dc.date.accessioned | 2021-07-27T08:09:16Z | - |
dc.date.available | 2021-07-27T08:09:16Z | - |
dc.date.issued | 2021 | - |
dc.identifier.citation | The 38th International Conference on Machine Learning (ICML), Virtual Conference, 18-24 July 2021. In Proceedings of Machine Learning Research (PMLR), v. 139: Proceedings of ICML 2021, p. 9934-9944 | - |
dc.identifier.issn | 2640-3498 | - |
dc.identifier.uri | http://hdl.handle.net/10722/301312 | - |
dc.description | Poster Session 2 | - |
dc.description.abstract | Object detection has recently achieved a breakthrough for removing the last one non-differentiable component in the pipeline, Non-Maximum Suppression (NMS), and building up an end-to-end system. However, what makes for its one-to-one prediction has not been well understood. In this paper, we first point out that one-to-one positive sample assignment is the key factor, while, one-to-many assignment in previous detectors causes redundant predictions in inference. Second, we surprisingly find that even training with one-to-one assignment, previous detectors still produce redundant predictions. We identify that classification cost in matching cost is the main ingredient: (1) previous detectors only consider location cost, (2) by additionally introducing classification cost, previous detectors immediately produce one-to-one prediction during inference. We introduce the concept of score gap to explore the effect of matching cost. Classification cost enlarges the score gap by choosing positive samples as those of highest score in the training iteration and reducing noisy positive samples brought by only location cost. Finally, we demonstrate the advantages of end-to-end object detection on crowded scenes. | - |
dc.language | eng | - |
dc.publisher | ML Research Press. The Journal's web site is located at http://proceedings.mlr.press/ | - |
dc.relation.ispartof | Proceedings of Machine Learning Research (PMLR) | - |
dc.relation.ispartof | The 38th International Conference on Machine Learning (ICML), 2021 | - |
dc.title | What Makes for End-to-End Object Detection? | - |
dc.type | Conference_Paper | - |
dc.identifier.email | Luo, P: pluo@hku.hk | - |
dc.identifier.authority | Luo, P=rp02575 | - |
dc.identifier.hkuros | 323746 | - |
dc.identifier.volume | 139: Proceedings of ICML 2021 | - |
dc.identifier.spage | 9934 | - |
dc.identifier.epage | 9944 | - |
dc.publisher.place | United States | - |