when validate the model，the gtbox num is wrong。

[caijillx]

After training with my own customized dataset, the number of gt-boxes in each category is smaller than the actual numbers. I don't know where the problem is. In addition, my dataset format is VOC.

Here are the output of the program.
[>>] 433/433, 6.6 task/s, elapsed: 66s, ETA: 0s
---------------iou_thr: 0.5---------------
2021-03-24 17:30:32,946 - mmdet - INFO -
+-----------+------+------+--------+-------+
| class | gts | dets | recall | ap |
+-----------+------+------+--------+-------+
| person | 853 | 4520 | 0.892 | 0.750 |
| bus | 175 | 924 | 0.749 | 0.452 |
| bicycle | 52 | 947 | 0.827 | 0.436 |
| car | 1820 | 4828 | 0.903 | 0.827 |
| motorbike | 101 | 965 | 0.871 | 0.621 |
+-----------+------+------+--------+-------+
| mAP | | | | 0.617 |
+-----------+------+------+--------+-------+

And this is the true gt-boxes nums
car:2579
bus:242
person:1113
motorbike:150
bicycle:68

[caijillx]

I think I found the reason, mmdetection ignored those tags with < difficulty > 1 < / difficulty> in the process of verification. So if it is feasible, how to modify the code to verify these ignored GT boxes?