search

sovit-123 / fasterrcnn-pytorch-training-pipeline

PyTorch Faster R-CNN Object Detection on Custom Dataset
MIT License
223 stars 75 forks source link

ValueError: y_max is less than or equal to y_min for bbox (tensor(0.9438), tensor(0.5500), tensor(0.9563), tensor(0.5500), tensor(26)) #13

Closed MheadHero closed 2 years ago

MheadHero commented 2 years ago

Hi, I was using the same dataset which split into train/valid/test. When I use train.py, it has no problem at all, but when I execute validate.py it comes out this error. You have any ideas about it? image

sovit-123 commented 2 years ago

Looks like dataset issue. The y_max should be larger than y_min obviously. But if you observe after normalization y_min = y_max = 0.5500. That's what is causing the error.

MheadHero commented 2 years ago

Looks like dataset issue. The y_max should be larger than y_min obviously. But if you observe after normalization y_min = y_max = 0.5500. That's what is causing the error.

Problem is I used the same dataset too in training, and training process went smooth but then when use validate.py it outputs error like this. So I think is not dataset issue. And if you google around this issue, there are similar issue in OTHER models saying that this is problem with albumentation. So if it was my dataset problem, it should be happened during training phase also, but it last till only validate.py only happens.

MheadHero commented 2 years ago

Please, can you help?

sovit-123 commented 2 years ago

Is it a public dataset that I can try out?

MheadHero commented 2 years ago

I can share you the dataset, actually you mind add me on discord, so we can directly communicate things about object detection? (if you don't shy haha)

dataset link: !curl -L "https://app.roboflow.com/ds/LsjVvA8COo?key=UA3zJoXRny" > roboflow.zip; unzip roboflow.zip; rm roboflow.zip

MheadHero commented 2 years ago

This dataset is resized into 320x320, and auto-orientation is applied. Exported in VOC format too

MheadHero commented 2 years ago

If you don't mind, you can add my discord (MheadHero, #2029), so maybe we can make an international friend right. Beside that, I also can directly ask you questions regarding object detection.

sovit-123 commented 2 years ago

Can you please check the friend request?

MheadHero commented 2 years ago

I have added you, and has hi to you, do you see it?