MMDetection 3.x分支有一个bug，就是训练的mask rcnn 导出模型部署小目标效果差~

[vansin]

[copy from wechat mmdet group3]
我发现一个MMDetection 3.x分支有一个bug，就是训练的mask rcnn 导出模型部署的时候，比较小的物体识别效果很差很差，差到没有眼看那种。
反而默认分支的MMDetection就没有这个问题。
训练精度都很好，直接用Python检测也很好，就是导出部署模型的时候就对比较小物体识别很差很差。我用一模一样的数据做了对比。

mmdeploy部署，或者自己写的libtorch都是一样

开始我是用MMDetection 默认分支，后面换成MMDetection 3.x分支，现在发现部署的时候比较小的物体识别很差很差，又换回来MMDetection 默认分支了。

MMDetection 3.x分支识别的效果真的没有眼看了

(上图为2.x版本)

（上图为 3.x 版本）

直接用Python是正常的，只要转换模型部署就出问题了。

同一台电脑，同一份数据，同一个环境

3.x最新分支，和默认最新分支

[1095788063]

训练精度是正常。就是导出模型的时候就识别很差了。

[vansin]

主要是部署有问题，锅首先会推到 mmdeploy 那边~

[vansin]

没有统计过，用肉眼都看出来不正常了。

肉眼不正常，也可以贴一张图

[1095788063]

我用这个代码也是一样有问题 libtorch(https://github.com/threeYANG/mmdetection_libtorch) 。
3.x 分支有问题，2.x 分支没有问题，这个就可以说明跟 mmdeploy没有关系，是 3.x 分支有问题了。同时 mmdeploy 也是一样有这个问题，那么说明 3.x 分支有问题。

[1095788063]

等下周，我做一个完整的说明，举证出来 3.x 分支有问题。

[1095788063]

您方便的时候测一下~

我找到原因了，是 3.x 的版本，部署的时候尺寸输入大小必须要和训练的尺寸大小一致才正常。如果部署的时候的输入尺寸大小和训练的尺寸大小不一致就会出现这个问题。同时 rtmdet-ins 我也测试也是这个问题。

[1095788063]

3.x 训练日志：
2023/01/12 14:05:22 - mmengine - INFO - Evaluating bbox...
2023/01/12 14:05:24 - mmengine - INFO -
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.903
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.984
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.980
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.843
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.913
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.939
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.939
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.939
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.882
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.948

2023/01/12 14:05:24 - mmengine - INFO -
计算每个类别的 Precision (AP) @[ IoU=0.50:0.95]
+------+---------------------------+---------------------------------+
| 类id | 类名称 | Precision (AP) @[IoU=0.50:0.95] |
+------+---------------------------+---------------------------------+
| 0 | HF-4CH | 0.973 |
| 1 | HF-LA | 0.815 |
| 2 | HF-LV | 0.894 |
| 3 | HF-RA | 0.78 |
| 4 | HF-RV | 0.879 |
| 5 | PP-Chest-Axial-Planes | 0.978 |
| 6 | PP-Chest-Axial-Planes-Cut | 0.998 |
+------+---------------------------+---------------------------------+
2023/01/12 14:05:24 - mmengine - INFO - bbox_mAP_copypaste: 0.903 0.984 0.980 -1.000 0.843 0.913
2023/01/12 14:05:24 - mmengine - INFO -
------------[bbox]计算验证结果耗时：1.367415189743042秒
2023/01/12 14:05:24 - mmengine - INFO - Evaluating segm...
2023/01/12 14:05:26 - mmengine - INFO -
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.881
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.983
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.970
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.773
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.905
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.914
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.914
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.914
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.830
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.927

2023/01/12 14:05:26 - mmengine - INFO -
计算每个类别的 Precision (AP) @[ IoU=0.50:0.95]
+------+---------------------------+---------------------------------+
| 类id | 类名称 | Precision (AP) @[IoU=0.50:0.95] |
+------+---------------------------+---------------------------------+
| 0 | HF-4CH | 0.976 |
| 1 | HF-LA | 0.778 |
| 2 | HF-LV | 0.86 |
| 3 | HF-RA | 0.726 |
| 4 | HF-RV | 0.851 |
| 5 | PP-Chest-Axial-Planes | 0.978 |
| 6 | PP-Chest-Axial-Planes-Cut | 0.996 |
+------+---------------------------+---------------------------------+
2023/01/12 14:05:26 - mmengine - INFO - segm_mAP_copypaste: 0.881 0.983 0.970 -1.000 0.773 0.905
2023/01/12 14:05:26 - mmengine - INFO -

[vansin]

3.x 的 Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.903 不是比 2.x 高吗？

哪里体现了小目标检测效果不好？ 都是 -1.000

[1095788063]

python 代码正常，导出模型部署的时候就不正常了。用mmdeploy 或者 libtorch(https://github.com/threeYANG/mmdetection_libtorch) 来部署的时候就出现识别效果很差的了。

[1095788063]

2.x 训练日志：

2023-01-13 18:12:30,644 - mmdet - INFO - Saving checkpoint at 50 epochs
2023-01-13 18:13:02,953 - mmdet - INFO - Evaluating bbox...
2023-01-13 18:13:04,278 - mmdet - INFO -
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.869
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.986
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.977
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.799
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.879
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.905
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.905
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.905
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.832
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.916

2023-01-13 18:13:04,279 - mmdet - INFO -
计算每个类别的 Precision (AP) @[ IoU=0.50:0.95]
+---------------------------+---------------------------------+
| 类别 | Precision (AP) @[IoU=0.50:0.95] |
+---------------------------+---------------------------------+
| HF-4CH | 0.954 |
| HF-LA | 0.750 |
| HF-LV | 0.854 |
| HF-RA | 0.748 |
| HF-RV | 0.825 |
| PP-Chest-Axial-Planes | 0.959 |
| PP-Chest-Axial-Planes-Cut | 0.989 |
+---------------------------+---------------------------------+
2023-01-13 18:13:04,279 - mmdet - INFO - Evaluating segm...
2023-01-13 18:13:05,849 - mmdet - INFO -
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.850
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.986
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.971
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.726
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.871
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.886
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.886
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.886
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.781
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.900

2023-01-13 18:13:05,850 - mmdet - INFO -
计算每个类别的 Precision (AP) @[ IoU=0.50:0.95]
+---------------------------+---------------------------------+
| 类别 | Precision (AP) @[IoU=0.50:0.95] |
+---------------------------+---------------------------------+
| HF-4CH | 0.958 |
| HF-LA | 0.713 |
| HF-LV | 0.826 |
| HF-RA | 0.701 |
| HF-RV | 0.809 |
| PP-Chest-Axial-Planes | 0.959 |
| PP-Chest-Axial-Planes-Cut | 0.984 |
+---------------------------+---------------------------------+

[1095788063]

3.x 训练验证过程图片

[1095788063]

2.x 训练验证过程图片

[1095788063]

3.x 和 2.x 的代码均为最新的代码

[1095788063]

我找到原因了，是 3.x 的版本，部署的时候尺寸输入大小必须要和训练的尺寸大小一致才正常。如果部署的时候的输入尺寸大小和训练的尺寸大小不一致就会出现这个问题。同时 rtmdet-ins 我也测试也是这个问题。

[ZwwWayne]

Thanks for your patience and report. We have noticed this issue and will fix it A.S.A.P. Please be noted that this might take some time because we are in the vacation of the Spring Festival.

[1095788063]

好的，我测试发现 rtmdet-ins 同样也有这个问题。修改输入尺寸，就识别效果变很差。一样的问题。