There are multiple bounding boxes on a single object.

[Ian-Work-AI]

The model works well in most cases, but sometimes it detects multiple bounding boxes for the same single object. The DETR paper says that the model will generate an independent bounding box for each individual target, so they didn't need NMS. Shouldn't the output keep only the highest scoring bounding box?

Thanks for your help!!

[PaulineTreyvaud]

I do have the same issue too. I obtain several perfectly overlapping boxes with different classes and even boxes with high IoU with the same label. Have you ever noticed that @lyuwenyu ?

[lyuwenyu]

I think you should choose score threshold carefully.

[David-dotcom666]

@lyuwenyu 应该不是分数阈值设置合理性的问题吧，会出现分数非常相近的几乎重叠的情况

[20231211]

@David-dotcom666 @PaulineTreyvaud 你好，我也遇到了相同的问题，请问你们找到解决方法了吗？

[PaulineTreyvaud]

@20231211 , the only way to mitigate the outcome is to consciously select confidence threshold (in my case, different for each class).

[David-dotcom666]

@David-dotcom666 @PaulineTreyvaud 你好，我也遇到了相同的问题，请问你们找到解决方法了吗？

调整阈值是一个方法，但是很多时候阈值会非常接近，自己加NMS也是一个，但是会考虑耗时。因为我的场景对这种偶尔多一个重叠没什么关系，我后面就没处理了，另外如果模型训练的比较好的时候这种情况出现的少。也可以问问作者有什么想法没。