search

lilanxiao / Rotated_IoU

Differentiable IoU of rotated bounding boxes using Pytorch
MIT License
416 stars 64 forks source link

a problem when using 3d-giou for regression training #47

Open Gretacyh opened 2 years ago

Gretacyh commented 2 years ago

Hello,There is a problem when using 3d-giou for regression training. When I use 3d-giou to train 3dssd on my own dataset(5 classes: car perdestrian headstock fulltrailer emptytrailer), it seems that giou implementation extremely harmful for categories with large aspect ratio.(for example: when training the valid mAP for trailer is always 0, however the mAP for other types is normal and progressively increase. ) Additionally, it can be normally training using iou/ciou/diou for regression. Last, hope for your reply.Thanks~

lilanxiao commented 2 years ago

Hi! For me, it looks like a technical issue. I've personally used 3D-GIoU for 3D detection in an indoor environment. Some classes also had large aspect ratios, e.g., bookshelves and cabinets. But I didn't see such a problem. I think some hyperparameters in your training might not be optimally chosen. According to my experience, the weighting factors of different sub-losses, e.g., for bounding box regression and semantic classification, have a significant impact on the training. Inappropriate weighting factors might lead to weird behaviors.

I don't know any details about your implementation so I might be wrong. But I would suggest tuning the hyperparameters to see if the problem persists.

Gretacyh commented 2 years ago

Thanks for your reply，I will check it later and tell your about the solvement if it does works.

半度微笑 @.***

------------------ Original ------------------ From: Lanxiao Li @.> Date: Tue,Aug 9,2022 9:40 PM To: lilanxiao/Rotated_IoU @.> Cc: Gretacyh @.>, Author @.> Subject: Re: [lilanxiao/Rotated_IoU] a problem when using 3d-giou forregression training (Issue #47)