2022-07-04-组会

Gliding Vertex

RSDet

四边形的检测还是挺有意义的，毕竟四边形相比旋转矩形的定位更加精确。

这两篇文章也让我知道了，要想考虑使用四边形检测，需要要考虑标签顺序问题和边界回归问题。

Others

最近使用batchsize = 8重跑了一次ReDet

5月底第一次用batchsize=1来跑

2022-05-31 11:41:46,926 - mmrotate - INFO - Saving checkpoint at 12 epochs
2022-05-31 13:22:35,774 - mmrotate - INFO - 
+--------------------+-------+--------+--------+-------+
| class              | gts   | dets   | recall | ap    |
+--------------------+-------+--------+--------+-------+
| plane              | 18788 | 23259  | 0.896  | 0.799 |
| baseball-diamond   | 1087  | 2135   | 0.724  | 0.651 |
| bridge             | 4181  | 4962   | 0.584  | 0.470 |
| ground-track-field | 733   | 1090   | 0.618  | 0.556 |
| small-vehicle      | 58868 | 110625 | 0.841  | 0.748 |
| large-vehicle      | 43075 | 74399  | 0.905  | 0.842 |
| ship               | 76153 | 88620  | 0.869  | 0.805 |
| tennis-court       | 5923  | 9230   | 0.937  | 0.904 |
| basketball-court   | 1180  | 2564   | 0.770  | 0.700 |
| storage-tank       | 13670 | 15183  | 0.674  | 0.622 |
| soccer-ball-field  | 827   | 2764   | 0.625  | 0.472 |
| roundabout         | 973   | 1973   | 0.623  | 0.544 |
| harbor             | 15468 | 22735  | 0.791  | 0.687 |
| swimming-pool      | 3836  | 8011   | 0.794  | 0.644 |
| helicopter         | 1189  | 1946   | 0.807  | 0.765 |
+--------------------+-------+--------+--------+-------+
| mAP                |       |        |        | 0.681 |
+--------------------+-------+--------+--------+-------+
2022-05-31 13:22:35,861 - mmrotate - INFO - Exp name: redet_re50_refpn_1x_dota_le90.py
2022-05-31 13:22:35,861 - mmrotate - INFO - Epoch(val) [12][12800]	mAP: 0.6807

7月初第二次用batchsize=8(ReDet论文中的同一batchsize)来跑，结果近似。。

好像Batch Normalization在这里并没有提高精度

2022-07-02 19:50:12,213 - mmrotate - INFO - Saving checkpoint at 12 epochs
2022-07-02 20:20:42,710 - mmrotate - INFO - 
+--------------------+-------+--------+--------+-------+
| class              | gts   | dets   | recall | ap    |
+--------------------+-------+--------+--------+-------+
| plane              | 18788 | 39421  | 0.931  | 0.892 |
| baseball-diamond   | 1087  | 5512   | 0.857  | 0.711 |
| bridge             | 4181  | 25338  | 0.636  | 0.473 |
| ground-track-field | 733   | 6700   | 0.782  | 0.552 |
| small-vehicle      | 58868 | 155702 | 0.807  | 0.703 |
| large-vehicle      | 43075 | 103456 | 0.882  | 0.766 |
| ship               | 76153 | 112373 | 0.839  | 0.790 |
| tennis-court       | 5923  | 13502  | 0.932  | 0.901 |
| basketball-court   | 1180  | 7586   | 0.714  | 0.615 |
| storage-tank       | 13670 | 33506  | 0.696  | 0.615 |
| soccer-ball-field  | 827   | 6799   | 0.674  | 0.454 |
| roundabout         | 973   | 8810   | 0.733  | 0.549 |
| harbor             | 15468 | 30952  | 0.773  | 0.658 |
| swimming-pool      | 3836  | 11873  | 0.810  | 0.688 |
| helicopter         | 1189  | 3929   | 0.866  | 0.786 |
+--------------------+-------+--------+--------+-------+
| mAP                |       |        |        | 0.677 |
+--------------------+-------+--------+--------+-------+
2022-07-02 20:20:42,713 - mmrotate - INFO - Exp name: redet_re50_refpn_1x_dota_le90.py
2022-07-02 20:20:42,713 - mmrotate - INFO - Epoch(val) [12][12800]	mAP: 0.6769