1 / 31

Anchor-free Detection 专题

Anchor-free Detection 专题. Lejian Ren 20190512. 为什么要有 anchor. 直接预测坐标难以训练 F(x) =(x1, y1, x2, y2). 早期的 anchor-free: DenseBox. FCN 形式,输出 5 通道( score + 4 个坐标变换) 对物体 size 差距较大的图像效果不好( FPN? ).

cherylg
Télécharger la présentation

Anchor-free Detection 专题

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Anchor-free Detection专题 Lejian Ren 20190512

  2. 为什么要有anchor • 直接预测坐标难以训练 • F(x) =(x1, y1, x2, y2)

  3. 早期的anchor-free: DenseBox • FCN形式,输出5通道(score + 4个坐标变换) • 对物体size差距较大的图像效果不好(FPN?) Huang, Lichao, et al. "Densebox: Unifying landmark localization with end to end object detection." arXiv preprint arXiv:1509.04874 (2015).

  4. Denet 预测每个像素是什么类型的corner DeNet Tychsen-Smith, Lachlan, and Lars Petersson. "Denet: Scalable real-time object detection with directed sparse sampling." Proceedings of the IEEE International Conference on Computer Vision. 2017.

  5. CornerNet Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  6. CornerNet • 为什么不用anchor • anchor多 • 超参多 • 为什么用corner • Center比较难定位,需要考虑和四个边的关系 • Corner更高效 Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  7. CornerNet Heatmap参考了key point Embedded是一个d-1的数 Offset用来减小从feature map到原图的影响(共享) Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  8. CornerNet • 预测corner • 高斯 • Focal loss Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  9. CornerNet • Corner pooling • 为什么需要:物体的corner往往不在物体里面 Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  10. CornerNet • Grouping Corner Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  11. CornerNet Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  12. CornerNet Law, Hei, and Jia Deng. "Cornernet: Detecting objects as paired keypoints." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

  13. CornerNet-Lite • 加速 • CornerNet-Saccade • 减小输入尺寸 • 利用attention减少需要预测的点 • CornerNet-Squeeze • 修改backbone Law, Hei, et al. "CornerNet-Lite: Efficient Keypoint Based Object Detection." arXiv preprint arXiv:1904.08900 (2019).

  14. CornerNet-Lite • CornerNet-Saccade Law, Hei, et al. "CornerNet-Lite: Efficient Keypoint Based Object Detection." arXiv preprint arXiv:1904.08900 (2019).

  15. CornerNet-Lite Law, Hei, et al. "CornerNet-Lite: Efficient Keypoint Based Object Detection." arXiv preprint arXiv:1904.08900 (2019).

  16. CenterNet-1 • 预测key point • 预测长宽 X. Zhou, D. Wang, and P. Krähenbühl, “Objects as Points,” 2019.

  17. CenterNet-1 • Loss functions X. Zhou, D. Wang, and P. Krähenbühl, “Objects as Points,” 2019.

  18. CenterNet-1 X. Zhou, D. Wang, and P. Krähenbühl, “Objects as Points,” 2019.

  19. CenterNet-2 • CenterNet的缺点 • 原文认为:只预测corner无法很好的catch物体本身的信息 • 我认为:embedding没有区分度 • 从预测corner变成预测corner+center • 只有预测的center在预测的corner内部,才是正确的预测 Duan, Kaiwen, et al. "CenterNet: Keypoint Triplets for Object Detection." arXiv preprint arXiv:1904.08189 (2019).

  20. CenterNet-2 Duan, Kaiwen, et al. "CenterNet: Keypoint Triplets for Object Detection." arXiv preprint arXiv:1904.08189 (2019).

  21. CenterNet-2 • 如何判断center在corner内部(判定区域的设定) • 太大,precision低 • 太小,recall低 • 自适应 Duan, Kaiwen, et al. "CenterNet: Keypoint Triplets for Object Detection." arXiv preprint arXiv:1904.08189 (2019).

  22. CenterNet-2 • Center pooling • Cascade corner pooling Duan, Kaiwen, et al. "CenterNet: Keypoint Triplets for Object Detection." arXiv preprint arXiv:1904.08189 (2019).

  23. CenterNet-2 Duan, Kaiwen, et al. "CenterNet: Keypoint Triplets for Object Detection." arXiv preprint arXiv:1904.08189 (2019).

  24. RepPoints(representative points) • motivation: • 物体空间形状的可以通过几个reppoints来表示,通过学多个reppoints,就可以用其外接矩形来表示物体的框 Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “RepPoints: Point Set Representation for Object Detection,” 2019.

  25. RepPoints 和deformable convolution 完美结合 Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “RepPoints: Point Set Representation for Object Detection,” 2019.

  26. RepPoints • Pipeline: • Learning RepPoints • RepPoints refinement • Converting RepPoints to bounding box • Min-max function • Partial min-max function • Moment-based function (reppoints均值和二阶矩) Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “RepPoints: Point Set Representation for Object Detection,” 2019.

  27. RepPoints Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “RepPoints: Point Set Representation for Object Detection,” 2019.

  28. RepPoints Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “RepPoints: Point Set Representation for Object Detection,” 2019.

  29. 一点看法 • 为什么object detection没有一开始就anchor-free? • 目前的anchor-free就是“anchor-free”吗? • Anchor-free可能的发展方向

More Related