WACV 2024 · Paper Collection
WACV 2024 includes relatively few directly related papers but has a strong application focus: redesigning keypoints, auditing acquisition-marker bias, controlling synthetic-data 3D consistency, and incorporating event cameras, active LEDs, and workspace lighting into estimation systems.
5
Papers included
4
Core collection
2
Active perception / robotic systems
1
Extended collection
We screened the titles of all 846 WACV 2024 papers on CVF Open Access, then checked candidate abstracts, outputs, and experimental subjects, including 4 core papers and 1 extended systems paper.
View the official WACV 2024 proceedings →Collection in One Sentence
KeyGNet learns keypoint layouts, while Marker Dataset Bias shows how markers and inpainting traces become pose shortcuts. Both highlight input representations and data-production processes as parts of the estimator.
P3D-Diffusion narrows synthetic appearance gaps while using edge conditioning to preserve controlled 3D viewpoints. Two other works directly control sensing and environments, using microsecond event streams or directional lighting for speed and fine geometry.
Research Trends
Trends summarize key methodological variables. Each paper belongs to one primary group, while tags retain cross-area attributes.
01 · Geometry Matters
KeyGNet constrains keypoint-voting distributions through a Wasserstein term and preserves geometric spread through dispersion, showing gains across SISO / MIMO pipelines beyond fixed box corners or FPS.
02 · Data Forensics
Stable geometry between cameras, marker boards, and objects makes background markers pose-correlated. Retaining markers, removing them by inpainting, or random augmentation changes learned cues and requires separate auditing, like label noise.
03 · Synthetic, but Geometric
P3D-Diffusion renders CAD at precise viewpoints before edge-guided diffusion style transfer. Realism, background diversity, and 3D consistency are treated as distinct objectives.
04 · Active Perception
Event cameras and active LEDs use frequency coding for fast correspondence. Directional lighting and photometric stereo use controlled reflectance for normals and edges. Both add hardware priors to obtain speed or precision difficult to achieve simultaneously with passive RGB.
Data Statistics
The 5 papers form three primary-contribution groups; active perception includes 1 core method and 1 extended system.
Examines how keypoint layout, marker appearance, and annotation noise affect final 6DoF pose, revisiting fixed heuristic keypoints and the assumption that training images merely carry object content.
Uses diffusion to improve synthetic-image realism while retaining controlled viewpoint labels, and reduces synthetic-to-real gaps through cross-domain 3D consistency.
Incorporates event cameras, active LEDs, and controlled directional lighting into estimator design for high-frequency, low-latency, or fine-surface-geometry robotic localization and manipulation.
4 + 1
4 papers primarily output object pose or category-level 3D orientation. Shape from Shading focuses on surface perception but includes pose / tracking experiments and is listed separately as an extension.
3 observation paradigms
Conventional RGB / RGB-D, event streams with active markers, and controlled multilight images reflect WACV's application and systems focus.
2 generalization problems
KeyGNet addresses geometric generalization across objects / estimators; P3D-Diffusion addresses synthetic-to-real and OOD generalization.
Recommended Reading
The CVF annual proceedings do not provide consistent Oral / Highlight labels for these papers. Recommendations are editorial selections and are not presented as official honors.
Uses LED frequency identities to establish 2D–3D correspondences, combining sensor timing, marker design, and PnP in a complete 3 kHz system.
Data diagnosisShows that acquisition markers can provide pose-correlated shortcuts, while simple inpainting can introduce new domain biases.
Priority readingKeyGNet learns geometric design variables usually fixed as box corners or FPS points, validating gains across three voting-based pipelines, SISO / MIMO modes, and seven BOP core datasets.
Priority readingSeparates image realism from preservation of renderer-controlled rotation labels, using CC3D for cross-domain feature consistency on neural-mesh vertices. Relevant to synthetic data and limited-label rotation estimation.
Further readingPrimarily studies surface normals and depth edges under active lighting, while demonstrating how these representations support known-model alignment and relative planar-motion estimation without CAD.
Paper List
Cards specify inputs, task boundaries, core / extended scope, and verifiable results. All five papers have coarse and deep notes with original figures and tables, accessible through site detail pages.
2 papers
Examines how keypoint layout, marker appearance, and annotation noise affect final 6DoF pose, revisiting fixed heuristic keypoints and the assumption that training images merely carry object content.
Janis Rosskamp · Rene Weller · Gabriel Zachmann
Task: Instance-level 6D pose; analyzes markers, marker-removal inpainting, and label errors in real data acquisition
Input: RGB + CAD / pose labels
Controlled experiments with GDR-Net, one tea-box object, and a YCB-V subset compare retained markers, inpainting, and random-marker augmentation, alongside pose-label noise. Turns marker-induced pose leakage and bias after marker removal into measurable data-audit questions.
Yangzheng Wu · Michael Greenspan
Task: Multi-object 6DoF pose of known objects; learns keypoint layouts for downstream voting / PnP
Input: RGB or RGB-D (depending on the downstream pose estimator) + CAD mesh
KeyGNet uses EdgeConv to learn shared keypoints from visible colored surfaces of training objects. A Wasserstein term aligns keypoint-voting distributions, while dispersion preserves geometric spread. Replacing fixed keypoints improves PVN3D ADD(S) by 16.4 points on LMO-MIMO and mean MIMO AR by 12.0 points across seven BOP datasets.
1 papers
Uses diffusion to improve synthetic-image realism while retaining controlled viewpoint labels, and reduces synthetic-to-real gaps through cross-domain 3D consistency.
Jiahao Yang · Wufei Ma · Angtian Wang · Xiaoding Yuan · Alan Yuille · Adam Kortylewski
Task: Category-level 3D orientation with few real labels and many automatically labeled synthetic images for in-domain and OOD generalization
Input: single RGB + category-level 3D mesh model
P3D-Diffusion renders randomly rotated CAD, improves appearance through Canny-edge-conditioned diffusion, and reduces contextual shortcuts using category-independent random backgrounds. CC3D links real images to neural-mesh vertices through pseudo-poses. On PASCAL3D+, 10% real rotation labels yields 86.7 / 62.4 accuracy at 30° / 10°; 50% yields 90.7 / 71.4. Outputs rotation only, without translation.
2 papers
Incorporates event cameras, active LEDs, and controlled directional lighting into estimator design for high-frequency, low-latency, or fine-surface-geometry robotic localization and manipulation.
Gerald Ebmer · Adam Loch · Minh Nhat Vu · Roberto Mecca · Germain Haessig · Christian Hartl-Nesic · Markus Vincze · Andreas Kugi
Task: 6DoF pose of a known-geometry active LED marker board relative to an event camera, for high-speed drone and robot relative localization
Input: event stream + active LED marker geometry
Identifies LEDs in event streams by blinking frequency and solves IPPE-PnP using known board geometry. Reports latency below 0.5 ms and 3 kHz output, with 34.5±16 mm translation and 0.74±0.15° rotation errors over 2.1–4.8 m. Demonstrates drone flight up to 10 m/s and approximately 10 m range.
Arkadeep Narayan Chaudhury · Leonid Keselman · Christopher G. Atkeson
Task: Surface perception, grasping, and deformation measurement in controlled workspaces; pose applications include known-model alignment and unknown-object relative planar tracking
Input: multi-illumination RGB images + calibrated lights / optional CAD
Captures images under six low-angle directional lights and one overhead light, recovering normals and depth edges through photometric stereo. With CAD, analysis-by-synthesis combines normals, edges, and silhouettes, reporting 0.80–5.50 px edge misalignment on ten objects. Without CAD, registers two integrated point clouds, with reported planar-error components reaching 3.28 mm, 2.40 mm, and 1.28°. Strongly depends on static controlled lighting, near-Lambertian white objects, and high visibility.
Scope & Deduplication
The collection centers on whether object pose is a primary output and separates related application work from the main set.
Although primarily active-light surface perception, it explicitly evaluates known-model alignment and unknown-object relative-motion registration, with quantitative errors and failure conditions. It is retained as an extended system outside the 4 core papers.
MagneticPillars and HDMNet address outdoor LiDAR registration; SphereCraft and CGAPoseNet+GCAN output camera poses; multiview / specular-object reconstruction does not primarily output object SE(3). Human, hand, animal pose, and 2D MOT are excluded under the same rules.
All five works use their official WACV 2024 publications as primary records. KeyGNet and Shape from Shading arXiv pages are public-version links, not additional papers. No duplicate official versions were found in other annual collections.
Reading Status
All five papers have standalone coarse and deep notes covering method structure, key experiments, and failure conditions, accompanied by original method diagrams, main results, ablations, or system experiments.
All titles and recommendation cards use stable site routes. Detail pages retain official papers, PDFs, projects, and code links for checking primary sources.
