Run train_ mutil.py about some problems encountered in custom datasets

[xiaoxiongmaoxuexi]

Hello, I have a problem. When I use my own data set for multi-target detection, the accuracy of training is 0. But objects can be recognized normally. Do you know why? Thank you for your answer. Is the label used for single target detection and multi-target detection。
The following phenomena appear in training. But the official data set training is normal.
2022-03-15 15:02:37 Testing cup...
2022-03-15 15:02:37 Number of test samples: 895
2022-03-15 15:03:06 Acc using 5 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 10 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 15 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 20 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 25 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 30 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 35 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 40 px 2D Projection = 0.00%
2022-03-15 15:03:06 Acc using 45 px 2D Projection = 0.67%
2022-03-15 15:03:06 Acc using 50 px 2D Projection = 1.12%
2022-03-15 15:03:06 -----------------------------------
2022-03-15 15:03:06 tensor to cuda : 0.000142
2022-03-15 15:03:06 predict : 0.003501
2022-03-15 15:03:06 get_region_boxes : 0.026212
2022-03-15 15:03:06 eval : 0.001130
2022-03-15 15:03:06 total : 0.030986
2022-03-15 15:03:06 -----------------------------------
2022-03-15 15:03:06 Testing sugar...
2022-03-15 15:03:06 Number of test samples: 480
2022-03-15 15:03:18 Acc using 5 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 10 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 15 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 20 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 25 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 30 px 2D Projection = 0.00%
2022-03-15 15:03:18 Acc using 35 px 2D Projection = 0.42%
2022-03-15 15:03:18 Acc using 40 px 2D Projection = 2.92%
2022-03-15 15:03:18 Acc using 45 px 2D Projection = 19.58%
2022-03-15 15:03:18 Acc using 50 px 2D Projection = 32.50%
2022-03-15 15:03:18 -----------------------------------
2022-03-15 15:03:18 tensor to cuda : 0.000140
2022-03-15 15:03:18 predict : 0.003520
2022-03-15 15:03:18 get_region_boxes : 0.019432
2022-03-15 15:03:18 eval : 0.001209
2022-03-15 15:03:18 total : 0.024300
2022-03-15 15:03:18 -----------------------------------
2022-03-15 15:03:18 Testing driller...
2022-03-15 15:03:19 Number of test samples: 1009
2022-03-15 15:03:51 Acc using 5 px 2D Projection = 0.00%
2022-03-15 15:03:51 Acc using 10 px 2D Projection = 0.00%
2022-03-15 15:03:51 Acc using 15 px 2D Projection = 0.00%
2022-03-15 15:03:51 Acc using 20 px 2D Projection = 0.10%
2022-03-15 15:03:51 Acc using 25 px 2D Projection = 0.20%
2022-03-15 15:03:51 Acc using 30 px 2D Projection = 1.09%
2022-03-15 15:03:51 Acc using 35 px 2D Projection = 3.57%
2022-03-15 15:03:51 Acc using 40 px 2D Projection = 10.70%
2022-03-15 15:03:51 Acc using 45 px 2D Projection = 21.21%
2022-03-15 15:03:51 Acc using 50 px 2D Projection = 36.17%
2022-03-15 15:03:51 -----------------------------------
2022-03-15 15:03:51 tensor to cuda : 0.000172
2022-03-15 15:03:51 predict : 0.005028
2022-03-15 15:03:51 get_region_boxes : 0.020104
2022-03-15 15:03:51 eval : 0.003032
2022-03-15 15:03:51 total : 0.028337
2022-03-15 15:03:51 -----------------------------------
2022-03-15 15:03:51 Testing duck...
2022-03-15 15:03:52 Number of test samples: 1065
2022-03-15 15:04:35 Acc using 5 px 2D Projection = 0.00%
2022-03-15 15:04:35 Acc using 10 px 2D Projection = 0.00%
2022-03-15 15:04:35 Acc using 15 px 2D Projection = 0.19%
2022-03-15 15:04:35 Acc using 20 px 2D Projection = 1.22%
2022-03-15 15:04:35 Acc using 25 px 2D Projection = 6.48%
2022-03-15 15:04:35 Acc using 30 px 2D Projection = 12.39%
2022-03-15 15:04:35 Acc using 35 px 2D Projection = 14.84%
2022-03-15 15:04:35 Acc using 40 px 2D Projection = 15.87%
2022-03-15 15:04:35 Acc using 45 px 2D Projection = 15.96%
2022-03-15 15:04:35 Acc using 50 px 2D Projection = 16.15%
2022-03-15 15:04:35 -----------------------------------
2022-03-15 15:04:35 tensor to cuda : 0.000231
2022-03-15 15:04:35 predict : 0.005087
2022-03-15 15:04:35 get_region_boxes : 0.032968
2022-03-15 15:04:35 eval : 0.002247
2022-03-15 15:04:35 total : 0.040533
2022-03-15 15:04:35 -----------------------------------
2022-03-15 15:04:35 Testing glue...
2022-03-15 15:04:35 Number of test samples: 1036
2022-03-15 15:05:07 Acc using 5 px 2D Projection = 0.00%
2022-03-15 15:05:07 Acc using 10 px 2D Projection = 0.00%
2022-03-15 15:05:07 Acc using 15 px 2D Projection = 0.10%
2022-03-15 15:05:07 Acc using 20 px 2D Projection = 1.25%
2022-03-15 15:05:07 Acc using 25 px 2D Projection = 4.83%
2022-03-15 15:05:07 Acc using 30 px 2D Projection = 9.27%
2022-03-15 15:05:07 Acc using 35 px 2D Projection = 14.38%
2022-03-15 15:05:07 Acc using 40 px 2D Projection = 19.98%
2022-03-15 15:05:07 Acc using 45 px 2D Projection = 25.00%
2022-03-15 15:05:07 Acc using 50 px 2D Projection = 28.67%

[xiaoxiongmaoxuexi]

The cup is a dataset I made myself. Thank you for your answer, thank you.

[xiaoxiongmaoxuexi]

Another problem is that when the trained model is visualized, it can be recognized in the context of and making data sets. When I am in other contexts, the recognition effect is very poor or even unrecognizable.

[zczczdc]

Hey, did you succeed in printing the bounding boxes for multiple objects in a single image? I try to run the source program valid.py and the visualization works, but the source program valid_muti.py cannot be displayed visually. Have you ever encountered such a problem? Thanks for your answer

[asd646942825]

Why is the glue so effective in your multi-target recognition? When I trained, I couldn't train the original author at all. Is your dataset loading slow?
Thank you for your reply