"Multiple detections" problem during inference

Hello,

I imported a pre-trained model and trained it with my own dataset with these commands for the import :

#!wget https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth checkpoint=torch.load("detr-r50-e632da11.pth",map_location='cpu') del checkpoint["model"]["class_embed.weight"] del checkpoint["model"]["class_embed.bias"] torch.save(checkpoint,"test_logo.pth")

The model works pretty well but the problem is that sometimes it detects multiple bounding boxes for the same object and I don’t know how to correct it … Here are some examples (for your understanding, the model is trained to detect logos like Airbus, or Peugeot) :

Shouldn’t it keep only the box with the higher score ? And so avoid the multiple detections ?

Does this problem comes from the fact that the num_queries is set to 100 in the downloaded pretrained model ? My dataset only contains up to 20 logos per image maximum, so maybe num_queries is too high and it forces the model to detect more objects than it should …? But if I change the num_queries when I make the training I have a dimension error so I don’t know how to change the num_queries for my model anyway… And if the “multiple detection” problem does not come from the num_queries parameter, could you please tell me how to solve it ?

Any help would be very much appreciated 🙏 Thank you all !

[fmassa]

Hi,

The multiple detections might be due to many things:

• maybe you need longer training – this depends on the cardinality_error that you have in your logs
• maybe you need a higher threshold? Currently DETR is calibrated in such a way that it requires a fairly high threshold for detections.

I think the num_queries being 100 is not necessarily the issue here, I think it should be fine.

[fmassa]

@hzphzp we don’t use it for mAP computation, and we have explained it in more detail in https://github.com/facebookresearch/detr/issues/201