WACV 2025 · Paper Collection
WACV 2025 shifts attention toward assets and priors before pose estimation: using one reference image, replacing CAD with reconstructions, avoiding synthetic-texture shortcuts, and executing transformations derived from generated goals in robotic systems.
4
Papers included
3
Core collection
2
Model assets / data design
1
Extended systems
We screened the titles of all 929 WACV 2025 main-conference papers on CVF Open Access, then checked candidate abstracts, full-text task definitions, outputs, and protocols, including 3 core papers and 1 extended system.
View the official WACV 2025 proceedings →Collection in One Sentence
Gen2SM fills missing reference viewpoints with generative priors but outputs relative rotation only. The reconstruction benchmark shows that standard reconstruction metrics and final pose AR are not equivalent.
Randomized Texturing induces shape bias during rendering for material robustness. PACA connects generated goals, object matching, and robot execution, while highlighting that a 6-DoF system differs from a full SE(3) object-pose benchmark.
Research Trends
Trends follow the variables each paper actually addresses, distinguishing rotation-only estimation, standard 6D pose, and object transformations inside robotic systems.
01 · One Reference
With one reference RGB image, Gen2SM denoises both reference and query toward discrete intermediate viewpoints and matches them, improving relative-rotation identifiability across large viewpoint gaps.
02 · Task-Aware Assets
Geometric completeness, PSNR, and visual appearance do not reliably predict pose AR. Testing a reconstructed model in MegaPose and FoundationPose establishes whether it can replace CAD.
03 · Shape over Texture
Randomized Texturing disrupts object–material associations so detectors and pose networks depend more on shape. Results also show that improvements are not consistent across all backbones, augmentations, and datasets.
04 · Pose in the Loop
PACA establishes object-level representations between generated images and real RGB-D scenes, then computes transformations for rearrangement. Evaluation emphasizes matching and task success rather than BOP-style rotation / translation errors.
Data Statistics
The 4 papers form three groups; the core includes rotation-only estimation, full 6D pose data / training studies, and downstream benchmarking.
Given one object reference RGB image and one query, uses generative priors to bridge large viewpoint gaps. Outputs relative 3D rotation without translation.
Evaluates reconstructed models inside FoundationPose / MegaPose and investigates rendering-time texture randomization, examining how CAD assets and synthetic appearance constrain 6D pose.
Generates target images from text, matches real RGB-D scenes, and executes object transformations. Pose is an intermediate robotic-system variable rather than the sole output of a standard pose benchmark.
1 + 2 + 1
1 single-reference relative-rotation paper, 2 model-asset / synthetic-training studies, and 1 extended robotic-rearrangement system.
2 main evaluation types
Gen2SM uses angular-threshold accuracy; the two standard 6D pose studies use BOP AR / VSD / MSSD / MSPD.
1 scope distinction
PACA parameterizes object transformations as (x,y,z,θ). Its 6-DoF designation also involves viewing perspective and the robotic system, rather than full object SE(3) output.
Recommended Reading
The CVF annual proceedings provide no consistent Oral / Highlight labels for these entries. Recommendations are editorial selections, not official honors.
Addresses large single-reference viewpoint gaps by denoising both reference and query toward intermediate views and matching noisy features, explicitly reporting accuracy, viewpoint-range, and runtime limitations.
Essential benchmark readingTests whether better-looking reconstruction improves pose, empirically distinguishing standard reconstruction metrics, rendering quality, reconstruction time, and BOP AR.
Representative engineering approachChanges rendering-time UV texture sampling alone, evaluating five detectors and two pose estimators while retaining results where configurations such as RetinaNet do not improve.
Further readingConnects object representations, generated goals, and transformations in a complete robotic loop. However, 6-DoF primarily describes the observation and rearrangement system, while object transformations use (x, y, z, θ), distinct from standard SE(3) pose benchmarks.
Paper List
Cards prioritize input resources, actual outputs, core / extended scope, and verifiable results. Official CVF links remain until site detail pages are complete.
1 papers
Given one object reference RGB image and one query, uses generative priors to bridge large viewpoint gaps. Outputs relative 3D rotation without translation.
Yujing Sun · Caiyi Sun · Yuan Liu · Yuexin Ma · Siu Ming Yiu
Task: One-reference–one-query relative 3D rotation for unseen objects without CAD, multiple references, or object-specific retraining
Input: single reference RGB + single query RGB
Gen2SM uses Zero-1-to-3 for bilateral generation and matching at discrete intermediate viewpoints, then recovers relative rotation through Monte-Carlo search and iterative refinement. Rotation accuracy at 15° / 30° is 55.32 / 82.14 on NAVI and 58.26 / 72.61 on GSO. A default query takes approximately 6.5 s search + 1.3 s refinement; the lightweight version takes 1.12 s. It estimates no translation and explicitly remains short of few-degree accuracy.
2 papers
Evaluates reconstructed models inside FoundationPose / MegaPose and investigates rendering-time texture randomization, examining how CAD assets and synthetic appearance constrain 6D pose.
Varun Burde · Assia Benbihi · Pavel Burget · Torsten Sattler
Task: Replaces YCB-V CAD with image-reconstructed models and evaluates downstream 6D pose using fixed general-purpose estimators
Input: calibrated multi-view RGB for reconstruction + RGB / RGB-D pose test images
Covers 21 YCB-V objects, evaluating COLMAP, RealityCapture, Nerfacto, VolSDF, MonoSDF, and BakedSDF reconstructions through MegaPose and FoundationPose. Even the best reconstructions retain an approximately 10–20% pose-quality gap to CAD. RealityCapture and Nerfacto offer strong accuracy–time tradeoffs; 75–100 surrounding images usually approach full-data performance, while fine-textured and reflective objects are harder.
Peter Hönig · Stefan Thalhammer · Jean-Baptiste Weibel · Matthias Hirschmanner · Markus Vincze
Task: Known-instance detection and 6D pose, emphasizing synthetic-to-real robustness for textureless and metallic industrial objects
Input: synthetic RGB + CAD meshes; RGB test images
During BlenderProc rendering, randomly samples 1,226 CC0 textures onto object surfaces to reduce reliance on fixed colors and materials. On T-LESS, GDR-Net BOP AR rises from 56.29 to 58.32 and Pix2Pose from 46.68 to 50.13; on ITODD, from 10.70 to 11.10 and 12.90 to 15.30, respectively. Gains depend on detector and augmentation settings, and incomplete ITODD ground truth limits conclusions about metal objects.
1 papers
Generates target images from text, matches real RGB-D scenes, and executes object transformations. Pose is an intermediate robotic-system variable rather than the sole output of a standard pose benchmark.
Shutong Jin · Ruiyu Wang · Kuangyi Chen · Florian T. Pokorny
Task: Text-goal-driven zero-shot tabletop rearrangement, estimating real-to-generated-goal object transformations for pick-and-place
Input: text prompt + wrist RGB-D + generated goal image
PACA builds object segmentations and features from Stable Diffusion denoising cross-attention and uses real-scene lines to control generated-goal perspective. ICP and depth alignment estimate object transformations. Mean matching accuracy over 450 image pairs is 86.72%; across 60 real robot trials, mean success in 3-DoF / 6-DoF scene settings is 68.73% / 65.56%. Success concerns relative-layout agreement with the generated goal, and transformations use (x,y,z,θ), not full three-axis rotation-error evaluation.
Scope & Deduplication
Pose in the WACV directory can refer to humans, hands, cameras, generation control, or object transformations; this collection distinguishes papers by their outputs.
PACA computes per-object transformations between real scenes and generated goals for pick-and-place, but its main task is scene rearrangement and success measures relative layout. Transformations include only (x,y,z,θ), so it is excluded from the three core pose papers.
Radiance Field-Based Pose Estimation outputs camera pose; Vision-Based Landing Guidance outputs camera-to-runway pose; Shift Equivariant Pose, STRIDE, and BioPose address human pose. Skip-and-Play is generic pose-controlled image generation without training or evaluating object-pose estimators. FineView provides camera calibration and multiview object data without an object-pose task.
All four works use official WACV 2025 papers as primary records. Gen2SM and Reconstruction → Pose Benchmark arXiv pages are public-version links, not additional papers. No duplicate official versions were found in other annual collections.
Reading Status
All four included papers have standalone coarse and deep notes with original method / result figures and site detail links. They clarify rotation-only outputs, CAD registration still required by the benchmark, shape bias's geometric and augmentation-tuning limits, and PACA's distinction between a 6-DoF system and full SE(3) object-pose output.
Collection-level checks cover metadata, differences between coarse and deep notes, original-figure counts, internal links, and build output. Deep notes are based on the paper's content rather than expanded abstracts or first-page screenshots.
