Submission name | |||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
Submission time (UTC) | Sept. 27, 2023, 11:20 a.m. | ||||||||||
User | sp9103 | ||||||||||
Task | Model-based 6D localization of seen objects | ||||||||||
Dataset | ITODD | ||||||||||
Description | |||||||||||
Evaluation scores |
|
User | sp9103 |
---|---|
Publication | https://arxiv.org/abs/2403.11510 |
Implementation | - |
Training image modalities | RGB-D |
Test image modalities | RGB |
Description | Submitted to: BOP Challenge 2023 Training data: MegaPose-GSO and MegaPose-ShapeNetCore Onboarding data: No Used 3D models: Default, CAD Notes: In this submission, GDRNPPDet_PBRReal [A] detections are used as the input to our pose estimation method. Our pose estimation method uses the coarse-to-fine strategy following the MegaPose [B] structure. A single model is used for all datasets. Our model was not trained on each BOP dataset, but only on the MegaPose dataset. We use the multi-hypotheses method proposed in the MegaPose [B]. The details are as follows: For each detection, we extract top-5 hypotheses from our coarse network, and each hypothesis is refined using the refinement network. The refined hypotheses are scored using the coarse network, and the best one is considered the output. Our coarse network is based on the MegaPose [B] coarse network. The main differences from the original MegaPose paper are as follows: To bypass the computation of rendering tons of images, we render fewer images and run the coarse network. Then we create a GMM with the top-k hypotheses and run the coarse network to the sampled hypotheses from the GMM. Our refinement network is based on the Shape-Constraint Recurrent Flow framework [C]. It estimates the flow from the rendered image to the input. The main differences from the original SCFlow paper are as follows: Our network estimates the visibility mask from the rendered image to the input. To train the visibility, we adopted the certainty estimation by classifying depth-consistency proposed in DKM [D]. We replace the pose regression network with the differentiable PnP solver to generalize the refiner well. Note that the inputs to our neural networks are the rgb images only, and the depth images in the training dataset are used to train the visibility mask. In this submission, each coarse hypothesis is refined 5 times. In the RGBD case, RANSAC-Kabsch is used for the depth refinement after estimating flow at every refinement step. [A] Liu et al.: https://github.com/shanice-l/gdrnpp_bop2022 List of contributors: Sungphill Moon (sungphill.moon@naverlabs.com), Hyeontae Son (son.ht@naverlabs.com) If you have any questions, feel free to contact us. |
Computer specifications | GPU A100; CPU Intel Xeon Gold6326@2.90G |