All Auto GT OD - Naifan Li's Bloghttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/All Auto GT OD | Naifan Li's BlogHugo -- gohugo.iozh-CNWed, 25 Feb 2026 10:38:44 +0800Auto GT ODhttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/draft-overview/Wed, 25 Feb 2026 10:38:44 +0800https://blog.omagiclee.com/posts/ai-data/auto-gt-od/draft-overview/Offboard 3D Object Detection from Point Cloud Sequences (CVPR 2021) Auto4D: Learning to Label 4D Objects from Sequential Point Clouds (2021)

Once Detected, Never Lost (CTRL) (ICCV 2023) MPPNet (OpenMMLab， 2022) DetZero (2023)r DetZero: Rethinking Offboard 3D Object Detection with Long-term Sequential Point Clouds (ICCV 2023) LablFormer: Object Trajectory Refinement for Offboard Perception (Waabi, 2023)

Fusion4DAL: Offline Multi-modal 3D Object Detection for 4D Auto-labeling SAM4D (2025)

]]>SAM4D: Segment Anything in Camera and LiDAR Streamshttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/sam4d/Thu, 26 Jun 2025 13:07:34 +0800https://blog.omagiclee.com/posts/ai-data/auto-gt-od/sam4d/ICCV 2025 arXiv 2506.21547 CN-ADLab/SAM4D Project

TL;DR

Motivations & Innovations

Approach

Model

Training Recipe

Data Recipe

Experiments

]]>DetZero: Rethinking Offboard 3D Object Detection with Long-term Sequential Point Cloudshttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/detzero/Fri, 09 Jun 2023 13:06:17 +0800https://blog.omagiclee.com/posts/ai-data/auto-gt-od/detzero/ arXiv

TL;DR

Motivations & Innovations

Approach

Model

Training Recipe

Data Recipe

Experiments

]]>3DAL: Offboard 3D Object Detection from Point Cloud Sequenceshttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/3dal/Mon, 08 Mar 2021 11:38:25 +0800https://blog.omagiclee.com/posts/ai-data/auto-gt-od/3dal/ Waymo LLC arXiv 2103.13612

TL;DR

Motivations & Innovations

Approach

Multi-frame 3D Object Detection

Multi-object Tracking

tracking-by-detection: using detector boxes for associations and Kalman filter for state updates.

Object Track Data Extraction

Object-centric Auto Labeling

Divide and conquer: motion state estimation

A linear classifier using a few heuristic features from the object track’s boxes can already achieve 99%+ motion state classification accuracy for vehicles.

]]>Auto4D: Learning to Label 4D Objects from Sequential Point Cloudshttps://blog.omagiclee.com/posts/ai-data/auto-gt-od/auto4d/Sun, 17 Jan 2021 11:56:49 +0800https://blog.omagiclee.com/posts/ai-data/auto-gt-od/auto4d/ arXiv 2101.06586

TL;DR

Motivations & Innovations

Approach

]]>