WACV 2026 · Paper Collection
The seven core WACV 2026 papers pursue complementary goals: better unseen-object accuracy and scalability under occlusion, deployable systems through multiview geometry, narrower outputs, or specialized hardware, and a reassessment by BOP-Distrib of whether one ground-truth pose adequately describes an image.
7
Core papers included
5
Full 6DoF methods
1
Rotation-only methods
1
Evaluation / benchmark papers
We screened the titles of all 831 WACV 2026 main-conference papers on CVF Open Access and checked candidate abstracts, outputs, and evaluation protocols. The collection includes 7 core papers, all labeled Poster in the official schedule.
Collection in One Sentence
WALDO and OPFormer advance unseen-object 6D pose through occlusion-related multistage error propagation and geometric encoding of foundation features. SDT-6D and KMOPS use multiview depth and stereo for scale and occlusion robustness.
FastPose-ViT retains full spacecraft 6DoF with projection correction, while Eff-GRot predicts rotation only to prioritize runtime. BOP-Distrib shows how hidden distinguishing parts can make asymmetric objects admit multiple poses.
Research Trends
Trends follow task assumptions and output variables, explicitly distinguishing full SE(3), rotation-only estimation, and pose-distribution benchmarks.
01 · Visibility
WALDO uses visibility-guided sampling, hypothesis retention, and refinement; BOP-Distrib defines equivalent poses from each image's visible surfaces. Both go beyond treating occlusion as uniform noise.
02 · Foundation + Geometry
OPFormer freezes DINOv2 but adds cross-layer aggregation, joint multitemplate encoding, and NOCS 3D positions before recovering correspondences.
03 · Multi-view Metric Cues
SDT-6D fuses multiview depth in sparse 3D voxels. KMOPS triangulates metric keypoints from stereo disparity without a depth camera, supporting transparent objects and multiple instances.
04 · Deliberate Efficiency
FastPose-ViT combines detection crops, closed-form geometry correction, FP16, and TensorRT. Eff-GRot omits translation and estimates unseen-object rotation in one forward pass.
Data Statistics
Seven papers form four complementary groups. Counts include only official WACV 2026 main-conference papers, excluding workshops, human pose, camera localization, and generic point-cloud registration.
WALDO and OPFormer incorporate object models, templates, and visible geometry for unseen-object inference. WALDO emphasizes occlusion and multi-hypothesis refinement; OPFormer emphasizes cross-template encoding and 2D–3D correspondence.
FastPose-ViT uses geometrically corrected direct regression for spacecraft edge deployment. Eff-GRot deliberately targets rotation only for real-time single-forward-pass inference.
SDT-6D builds sparse 3D voxels from multiview depth, while KMOPS triangulates keypoints from calibrated stereo. Both use multiview geometry to reduce occlusion and scale ambiguity.
BOP-Distrib derives per-image pose distributions from currently visible surfaces instead of one ground-truth pose, revisiting rankings of single-solution and multi-solution methods.
6 + 1
6 method papers and 1 paper on per-image pose distributions and new evaluation protocols.
5 / 1 / 1
5 output full 6DoF; 1 outputs rotation only; 1 revises ground truth and metrics rather than introducing an estimator.
7 Posters
The official Accepted Papers schedule lists all seven as Posters. Site recommendations are not Oral, Highlight, or Award labels.
Recommended Reading
All seven are official Posters. Recommendations below are editorial selections with stated reasons, rather than conference honors.
Jointly redesigns sampling, multi-hypothesis inference, iterative refinement, occlusion augmentation, and evaluation weighting to examine how occlusion propagates through multistage pipelines.
Representative engineering approachCombines multiview depth fusion, high-resolution sparse voxels, staged heatmaps, and per-voxel voting in one industrial pipeline, relevant to IPD, MV-YCB, and bin-picking deployment.
Representative efficient approachConnects task definition, projection correction, direct regression, and Jetson TensorRT deployment, explicitly distinguishing 75 ms sequential latency from 33 FPS concurrent throughput.
Priority readingJointly encodes multiple templates and NOCS 3D positions before decoding semidense 2D–3D correspondences, offering both CAD and NeRF onboarding routes.
Representative efficient approachRestricts output to rotation to test reference-based inference without correspondence, PnP, or iterative optimization, illustrating accuracy–latency–task-completeness tradeoffs.
Essential benchmark readingShows that global object symmetry alone cannot define equivalent poses for one image: viewpoint, self-occlusion, and external occlusion can make asymmetric objects visually ambiguous.
Representative engineering approachDecodes instances and keypoints in parallel across full stereo scenes, triangulates metric 3D keypoints from disparity, and obtains rotation, translation, and dimensions by closed-form registration.
Paper List
Cards clarify input priors, actual outputs, key mechanisms, and metric definitions. Links switch to site details when complete.
Unseen objects and model priors
WALDO and OPFormer incorporate object models, templates, and visible geometry for unseen-object inference. WALDO emphasizes occlusion and multi-hypothesis refinement; OPFormer emphasizes cross-template encoding and 2D–3D correspondence.
Sajjad Pakdamansavoji · Yintao Ma · Amir Rasouli · Tongtong Cao
WALDO concentrates computation through visibility-driven nonuniform dense sampling, retains multiple confidence-ranked initial poses, and refines them iteratively, with occlusion-targeted augmentation and visibility-weighted evaluation. Reports accuracy gains over 5% on ICBIN and over 2% on BOP, with approximately 3× faster inference. Interpret these alongside its revised visibility statistics and full experimental tables.
Artem Moroz · Vít Zeman · Martin Mikšík · Elizaveta Isianova · Miroslav David · Pavel Burget · Varun Burde
OPFormer detects objects with CNOS, then establishes semidense 2D–3D correspondences through frozen DINOv2, cross-layer weight adapters, a template encoder with NOCS 3D rotary positional encoding, and a cross-attention decoder. PnP recovers 6D pose. Templates come from CAD or iNGP trained on multiview images. Reported timings are under 1 minute/object for onboarding, under 0.5 s/image for detection, and under 0.05 s/instance for pose.
Efficient and specialized pose
FastPose-ViT uses geometrically corrected direct regression for spacecraft edge deployment. Eff-GRot deliberately targets rotation only for real-time single-forward-pass inference.
Pierre Ancey · Andrew Price · Saqib Javed · Mathieu Salzmann
FastPose-ViT directly regresses translation and apparent rotation from detection crops, then recovers full-image true pose through closed-form projection geometry without PnP. Leads among non-PnP methods on SPEED / SPEED+ and deploys on Jetson Orin Nano: approximately 75 ms/frame sequential end-to-end latency and up to 33 FPS pipelined throughput. The latter does not imply 30 ms single-frame latency.
Fanis Mathioulakis · Gorjan Radevski · Tinne Tuytelaars
Eff-GRot extracts global query and reference features through a frozen image encoder and passes reference rotation embeddings plus a query mask token to a transformer for single-pass rotation regression. Requires no CAD, object- or category-specific training, explicit correspondence, PnP, or refinement. Output is SO(3) rotation only, so runtime claims do not directly extend to complete 6D pipelines with translation.
Multiview metric pose
SDT-6D builds sparse 3D voxels from multiview depth, while KMOPS triangulates keypoints from calibrated stereo. Both use multiview geometry to reduce occlusion and scale ambiguity.
Nico Leuze · Maximilian Hoh · Samed Doğan · Nicolas R.-Peña · Alfred Schoettl
SDT-6D fuses multiview depth into fine-grained point clouds or sparse TSDFs. Staged heatmaps transfer scene-adaptive attention priors across resolutions; a density-aware sparse transformer handles occlusion and nonuniform 3D sampling; per-voxel voting outputs multiple object poses jointly. Evaluates on IPD and MV-YCB, emphasizing fully sparse high-resolution computation and competitive dense-bin performance.
Ying-Kun Wu · Yi Shen · Tzuhsuan Huang · I-Sheng Fang · Jun-Cheng Chen
KMOPS jointly predicts multiple object classes, 2D keypoints, disparities, and visibility using a shared RT-DETR encoder and grouped transformer decoders. Calibrated stereo triangulation yields metric 3D points; weighted SVD aligns canonical keypoints to recover rotation, translation, and dimensions. Extrinsics are not network inputs but calibration remains necessary for triangulation postprocessing. Evaluates on transparent-object TOD and large-scale StereOBJ-1M.
Visual ambiguity and evaluation
BOP-Distrib derives per-image pose distributions from currently visible surfaces instead of one ground-truth pose, revisiting rankings of single-solution and multi-solution methods.
Boris Meden · Asma Brazi · Fabrice Mayran de Chamisso · Steve Bourgeois · Vincent Lepetit
BOP-Distrib automatically generates per-image nonparametric multimodal 6D pose ground-truth distributions for T-LESS and YCB-V, incorporating visible-surface ambiguity into evaluation. New annotations substantially alter single-pose rankings. Pose-space precision / recall for estimators outputting a few modes enables the first quantitative comparison of multi-solution pose methods on real images.
Scope & Exclusions
Pose, rotation, registration, or tracking in a title does not automatically indicate object 6D pose estimation.
Detail Page Progress
Each detail page will provide coarse and deep notes, source-checked numbers, and 2–4 explanatory method diagrams, result tables, or failure cases. Paper cards switch to site details once these are complete.
