search

MakiseChris666 / MVXNet-Makise

An unofficial implementation of MVXNet
4 stars 0 forks source link

Thanks for your code, bro. #2

Open swaggywilliam opened 1 year ago

swaggywilliam commented 1 year ago

Have you tried running the MVXNet reproduction in mmdetection3d? Which one performs better, your reproduction or his? Also, the original MVXNet paper reports mAP 3D(IoU=0.7) is 70+, while the mmdetection3d reproduction only achieves scores in the 60+. Could it be due to different evaluation metrics being used? It's a question confused me a lot.

MakiseChris666 commented 1 year ago

Honestly my code is just a fundamental implement of MVX-Net and the code cannot work perfectly now, so obviously mmdet3d works better. For the inferior result produced by mmdet3d, I think that's because the original MVX-Net paper only used cars as targets. Actually it's hard for 3d detection models to detect small objects like pedestrians and cyclists, and mmdet3d's results just average the mAPs of cars, pedestrians and cyclists. You can find that mmdet3d could reach 90+ mAP on car detection, so it just works well. You can read the paper "Exploring Data Augmentation for Multi-Modality 3D Object Detection" to compare the performance between the original MVX-Net and mmdet3d's MVX-Net.