Need for NMS

[prashant-dn]

HI,
I realise that the model is based on DETR and uses the non-NMS approach. But after training my model and inferring at 0.3 detection threshold, I get multiple boxes.

But if I increase the threshold to 0.5, I do see fewer boxes but there are multiple boxes still left.
Increasing the threshold further only impacts other classes and I get overall lower detection boxes.

Am I missing something? What is the correct way to tackle this for such non-NMS based models?

[amirbakhtiary23]

No you are not missing anything. You just need to train more (if you have the enough data of course).
But that aside, you can do NMS too and since its only around 300 boxes it won't show any significant fps drop.

PS: Im not the author, just some random enthusiast

[prashant-dn]

Hey Amir, I can confirm that I have sufficient data and also the losses do look saturated. So training further makes doesn't seem like an option.

No you are not missing anything. You just need to train more (if you have the enough data of course). But that aside, you can do NMS too and since its only around 300 boxes it won't show any significant fps drop.

PS: Im not the author, just some random enthusiast

[amirbakhtiary23]

Are you fine tuning or training from scratch? if not fine tuning, do that instead. Moreover, Just because the loss becomes saturated it doesnt mean the model have converged well enough. Maybe you need to a bit of hyper parameter tuning as well.
Also, I dont know how many steps have you trained your model . I can see an 80K on the graphs but I dont see the batch size nor the dataset size.
Provide as much information as possible so I can help.