3D detection

Stereo 3D detection

LIGAStereo

pdf code

这篇paper是近期来性能第一个超过三/四年前点云检测的双目检测网络。性能上全面超越双目此前的网络。其主要思想是使用类似于 DSGN的网络， 但是于此同时，另外训练一个点云检测网络，两个网络输出受相同的监督损失训练，此外双目的cost volume分支在最后一层还会受点云最后一层的知识蒸馏训练。

 visualDet3D

include: Mono3D and Stereo3D

video effect: https://www.bilibili.com/video/BV1E7411u7Lv/?spm_id_from=333.788.recommend_more_video.3

github:  https://github.com/Owen-Liuyuxuan/visualDet3D

This repo contains the official implementation of 2021 RAL & ICRA paper Ground-aware Monocular 3D Object Detection for Autonomous Driving. Arxiv Page. Pretrained model can be found at release pages.

Also the official implementation of 2021 ICRA paper YOLOStereo3D: A Step Back to 2D for Efficient Stereo 3D Detection. Pretrained model can be found at release pages.

We further incorperate an Unofficial re-implementation of Monocular 3D Detection with Geometric Constraints Embedding and Semi-supervised Training (KM3D) as a reference on how to integrate with other frameworks. (Notice that the codes are from the originally official repo, and we DO NOT guarantee a complete re-implementation).

Update (2021.07.02): We provide an Unofficial re-implementation of Objects are Different: Flexible Monocular 3D Object Detection (MonoFlex) with few additional codes, based on the KM3D structure. Many of the core codes are from original official repo. We did not implement the edge merge operation and the corner loss, but we manage to maintain most of the performance based on the proposed depth fusion methods(validation AP reaches 15%).

Update (2021.12.11): We provide an Unofficial re-implmentation of Digging Into Output Representation For Monocular 3D Object Detection (Digging_M3D) to introduce an simple but important numerical trick to significantly improve the KITTI mAP scores and make a significant change to the KITTI leaderboard. Details can be found in the paper. At the time of the open-source, the paper has not been officially published, and we will keep up with the update of the paper.