search

Megvii-BaseDetection / YOLOX

YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/
Apache License 2.0
9.19k stars 2.17k forks source link

Blurred images after data augmentation #565

Open pransen opened 2 years ago

pransen commented 2 years ago

I am trying to fine-tune the model on my own dataset.

However, while looking into the images (for initial few iterations) after augmentation has been applied, I find the images to be quite blurry. Is this expected ?

Joker316701882 commented 2 years ago

@pransen Yes. It's normal. That's why we close the augmentation for the last 15 epochs.

pransen commented 2 years ago

Is it also normal for the images in the mosaic to overlap ? It's as if the images are blended together similar to cv2.addWeighted().

Joker316701882 commented 2 years ago

@pransen The overlap you mentioned is caused by Mixup. It is also closed for the last 15 epochs.

cena001plus commented 2 years ago

@pransen The overlap you mentioned is caused by Mixup. It is also closed for the last 15 epochs.

@Joker316701882 the next epoch is to turn off the enhancement, but the program is interrupted at this epoch. This problem consistently troubles me. Do you know why?

This problem occurs every time bboxes_preds_per_image is a null value. I want to know whether it is normal to have a null value, and whether there is a problem with the exception handling mechanism. Put the code that is prone to exceptions into the try?