Profile
Back to NewsBack
GitHub Trending 3 min
Reader Mode
WangYuLin-SEU/HCCEPose: [ICCV'25 Highlight] HccePose (BF): Predicting Front & Back Surfaces to Construct Ultra-Dense 2D-3D Correspondences for Pose Estimation

WangYuLin-SEU/HCCEPose: [ICCV'25 Highlight] HccePose (BF): Predicting Front & Back Surfaces to Construct Ultra-Dense 2D-3D Correspondences for Pose Estimation

HCCEPose

!HCCEPose

English · 中文

From image pixels to an object’s pose in 3D. HccePose(BF) predicts coordinates on both front and back surfaces of known objects and estimates object-to-camera rotation and translation from ultra-dense 2D–3D correspondences.

Find your starting point

| Goal | Documentation | | --- | --- | | Run pretrained inference | Installation → Models and inputs → First prediction | | Process images or video | RGB images · Video frames | | Estimate or refine with depth | WAPR · MegaPose · Depth comparison | | Train new objects | Meshes and symmetry → PBR rendering → Detector → Surface labels → Pose training | | Integrate into an application | Tester · Results and coordinates · Refinement · Data loading · Utilities | | Accelerate, evaluate or explore | ONNX/TensorRT · BOP evaluation · Public changelog · Project structure |

Get started

The project provides RGB inference, WAPR integration, and an optional interface to an external MegaPose installation. Prepare matching detector and object pose weights, object meshes, camera intrinsics, and input images, then run an inference example. See installation, resource conventions and your first prediction.

python examples/inference/predict_image.py

Run this command from the project root after editing the resource paths and object IDs at the end of the entry script. Outputs are saved to outputs/rgb/.

WAPR integration

Install WAPR separately. The example restores instance masks predicted by HccePose for the current image to the full frame, then estimates poses using registered depth, camera intrinsics, and each object’s metric mesh and diameter. Each instance produces pose_4x4_m (object to camera, translation in meters) and a score_6d computed for that input. Input instance order is preserved. This API estimates poses from masks without HccePose pose seeds. See WAPR usage.

Optional score filtering and per-object CAD rendered-mask NMS are provided in the postprocessing tutorial and examples/inference/postprocess_wapr.py. The raw inference JSON is kept separately.

Method

!HccePose(BF) paper figure

HccePose(BF) network architecture and pose estimation pipeline. See the method and results contract.

New objects and the complete training workflow

Public upstream includes mesh centering, rotational symmetry, BOP metadata, PBR rendering, YOLO training, front/back labels, per-object pose training and BOP evaluation scripts. Bilingual docs retain stage-by-stage procedures, public parameters, output layouts and example figures. Stage scripts remain in the repository root; for the original published revision, follow Installation; use this project’s examples/inference/ for application inference.

See BOP evaluation for public weights, historical results and leaderboard images. Those historical results describe the original published settings, rather than other data, models or configurations.

Documentation website

Project website · Documentation · All projects

python -m pip install -r docs/requirements.txt
python docs/build.py
python -m http.server 8768 --directory pages

Homepage: http://127.0.0.1:8768/; docs: http://127.0.0.1:8768/docs/index.html?lang=en.

See project structure, LICENSE and public sources. Third-party components retain their own licenses.

HccePose(BF), ICCV 2025

Citation

@InProceedings{HccePose_BF,
    author    = {Wang, Yulin and Hu, Mengting and Li, Hongli and Luo, Chen},
    title     = {HccePose(BF): Predicting Front \& Back Surfaces to Construct Ultra-Dense 2D-3D Correspondences for Pose Estimation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {7166-7175}
}
Chat with me