Method: OfficialDet-PFA-Mixpbr-RGB-D

User CVIA_Lab
Publication Yinlin Hu et, at: Perspective Flow Aggregation for Data-Limited 6D Object Pose Estimation, ECCV, 2022
Implementation
Training image modalities RGB
Test image modalities RGB-D
Description

Training data: Official PBR + Real (if possible).

Used 3D models: Official ones.

Object detection: Official detection results

Pose estimation: WDR

Pose refinement: PFA

Setting: Single model per dataset

We use the standard initialization-and-refinement framework, and train a single model for all objects in the same dataset. We use the official detection results as the preprocessing, WDR-Pose as the pose initialization, and PFA as our refinement.

All the mentioned models are trained purely on RGB data, and we only use the depth image during the inference to get the final pose in RGBD track by RANSAC-Kabsch.

Most of the framework is the same as the version we submitted last year .

The main differences include:

using the official detection;optimized batch size, training time, and other hyper-parameters;better Ransac-Kabsch

List of contributors:

Xidian University: Xinyao Fan, Fengda Hao, Yang Hai, Jiaojiao Li, Rui Song

EPFL: Haixin Shi, Mathieu Salzmann

Magic Leap: David Ferstl, Yinlin Hu

Computer specifications NVIDIA-3090,Intel xeon 4310

Public submissions

Date Submission name Dataset
2023-09-25 15:56 - LM-O
2023-09-25 15:56 - T-LESS
2023-09-25 15:57 - IC-BIN
2023-09-25 15:57 - HB
2023-09-25 15:59 - YCB-V
2023-09-25 15:59 - TUD-L
2023-09-25 16:20 - ITODD