hedonismbot24/foundationpose-6d-pose-estimation

🎯 FoundationPose — 6D Object Pose Estimation

[FoundationPose](https://nvlabs.github.io/FoundationPose/) by NVIDIA is the #1 method on the BOP Challenge 2024 benchmark for model-based 6D localization of unseen objects.

Features

• —Model-based initialization — Upload a 3D mesh (.obj/.stl/.ply) of your object
• —Automatic masking — SlimSAM-powered object segmentation
• —6D pose estimation — Full position (XYZ) + orientation (quaternion/Euler)
• —3D visualization — Projected coordinate axes and bounding box overlay

How to Use

1. 1.Initialize: Go to tab ①, upload a 3D mesh file and click "Initialize Object"
2. 2.Estimate: Go to tab ②, upload a query RGB image (+ optional depth) and click "Estimate Pose"
3. 3.View: See the pose axes overlaid on your image and the numeric pose output

Architecture

FoundationPose uses a two-stage pipeline:

1. 1.42 coarse pose hypotheses from uniformly sampled viewpoints
2. 2.Transformer-based refinement (ResNet-34 + 4-head attention)
3. 3.Contrastive ranking (InfoNCE) to select the best pose

Trained on 1.2M synthetic images from Objaverse — zero real-world training data needed.

BOP Challenge 2024 Results

Links

• —📄 Paper (CVPR 2024)
• —🌐 Project Page
• —💻 GitHub
• —🏋️ Model Weights
• —🏆 BOP Challenge