Skip to content
Computational Pathology Reading Roadmap

Video, 3D, and embodied perception

Images freeze one projection at one time. Video adds temporal sampling and identity; multiview vision adds camera geometry; embodied perception adds actions that change what can be observed. A model can perform well through static appearance while failing to use motion, geometry, or intervention, so each source of evidence needs a counterfactual test.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates.

Individual ICLR oral-paper studies retain their 19 September 2026 source-check dates; the later discovery audit does not re-date that evidence.

ECCV 2026 conference additions and metadata checked through 2026-09-12; the broader literature audit boundary above is unchanged.

Video world models and feed-forward geometry systems are integrated into temporal prediction and multiview reconstruction, not collected in dated sections.

Prerequisites and objectives

Section titled “Prerequisites and objectives”

Complete image recognition and the tensor and attention material in representation learning. You should be able to:

  • track space, time, view, point, and camera axes through a model;
  • distinguish motion evidence from repeated static appearance;
  • derive projection, epipolar, triangulation, and pose relationships;
  • compare points, voxels, meshes, radiance fields, and Gaussian scene primitives;
  • separate visual understanding, future prediction, video generation, and action planning;
  • select temporal, tracking, geometry, rendering, and control metrics correctly.

Video is a sampled spatiotemporal signal

Section titled “Video is a sampled spatiotemporal signal”

A clip x ∈ ℝ^(T × H × W × C) is defined by frame rate, exposure, clip duration, sampling stride, crop, and number of train/test views. Changing these changes the evidence.

Three common designs are:

  1. extract spatial features per frame and fuse them over time;
  2. use 3D convolution over space–time volumes;
  3. tokenize patches or tubes and apply factorized, local, or global attention.
Stage Primary reading · year · venue/status Read it for
V1 Two-Stream Convolutional Networks — 2014 · NeurIPS, peer reviewed separate appearance and optical-flow pathways
V2 Learning Spatiotemporal Features with 3D Convolutional Networks — 2015 · ICCV, peer reviewed homogeneous 3D convolution and clip descriptors
V3 Quo Vadis, Action Recognition? — 2017 · CVPR, peer reviewed inflating image filters into I3D and the value of large video pretraining
V4 SlowFast Networks — 2019 · ICCV, peer reviewed low-rate semantics plus a lightweight high-rate motion pathway
V5 TimeSformer — 2021 · ICML, peer reviewed full versus divided space–time attention
V6 Video Swin Transformer — 2022 · CVPR, peer reviewed local spatiotemporal windows and image-to-video transfer

Always include a single-frame and temporally shuffled baseline. If performance is unchanged, the model may be recognizing scenes or objects rather than actions.

Tracking and persistent identity

Section titled “Tracking and persistent identity”

Tracking estimates object state and identity through time. Detection quality, association, occlusion, re-identification, camera motion, and track management all affect the result.

An optical-flow estimator predicts a field u(p) such that brightness or features at p correspond to p + u(p) in the next frame. Tracking-by-detection instead associates ranked detections, often using motion and appearance costs.

Stage Primary reading · year · venue/status Read it for
T1 RAFT — 2020 · ECCV, peer reviewed all-pairs correlation volumes and recurrent dense-flow refinement
T2 ByteTrack — 2022 · ECCV, peer reviewed associating low-score detections rather than discarding them before tracking
T3 U²Flow — 2026 · CVPR, peer reviewed uncertainty-aware unsupervised flow; separate uncertainty calibration from endpoint error
T4 Adaptive Capacity Autoregressive Visual Tracking — 2026 · CVPR, peer reviewed adapting inference capacity to temporal difficulty while auditing latency and drift

Measure:

  • detection AP independently of association;
  • HOTA or a comparable joint detection/association metric;
  • IDF1 and identity switches;
  • mostly tracked/lost trajectories and fragmentation;
  • latency, causal versus offline access, and memory growth.

Do not report MOTA alone as a complete identity analysis, and do not compare an offline tracker using future frames with a causal streaming model as if their evidence were equal.

Video self-supervision and world prediction

Section titled “Video self-supervision and world prediction”

Video supervision can reconstruct pixels, predict masked tubes, align text, predict teacher features, or learn latent dynamics. Prediction target and temporal horizon determine what information is retained.

Branch Primary reading · year · venue/status Scientific boundary
masked pixels VideoMAE — 2022 · NeurIPS, peer reviewed tube masking and high mask ratios exploit temporal redundancy
latent feature prediction V-JEPA — 2024 · TMLR, peer reviewed predict target-region features without pixels, labels, text, or negatives
image/video prompts SAM 2 — 2025 · ICLR, peer reviewed streaming memory for prompted segmentation, not general physical reasoning
prediction and planning V-JEPA 2 — 2025 · arXiv technical report action-free internet-video pretraining plus limited action-conditioned data; treat planning claims separately from recognition
world-oriented masked learning InternVideo-Next — 2026 · CVPR, peer reviewed encoder–predictor–decoder separation and semantically guided latent dynamics

A predictive representation is not automatically a controllable world model. Evaluate counterfactual dynamics, multi-step drift, uncertainty, interventions, and planning success—not only video classification.

Camera geometry

Section titled “Camera geometry”

For a calibrated camera,

x̃ ∼ K[R | t]X̃

For two views, corresponding normalized image points satisfy an epipolar constraint x̃′ᵀEx̃ = 0. Triangulation estimates X from camera rays; bundle adjustment jointly refines camera and point parameters by minimizing reprojection error.

The classical pipeline—features, matches, robust geometry, triangulation, bundle adjustment, and multiview stereo—remains the diagnostic reference for learned 3D systems.

Stage Primary reading · year · venue/status Read it for
G1 Photo Tourism — 2006 · ACM SIGGRAPH, peer reviewed large unordered photo collections, structure from motion, and scene browsing
G2 ORB-SLAM — 2015 · IEEE T-RO, peer reviewed one ORB feature vocabulary across tracking, local mapping, relocalization, and loop closing; three parallel threads plus keyframe/map-point culling
G3 Structure-from-Motion Revisited — 2016 · CVPR, peer reviewed robust incremental SfM and the COLMAP pipeline
G4 PointNet — 2017 · CVPR, peer reviewed permutation-invariant point-set learning and its locality limitation
G5 Vision Transformers for Dense Prediction — 2021 · ICCV, peer reviewed reassembling transformer tokens for monocular depth and dense outputs

Depth is scale-ambiguous without additional information. State whether depth is metric, scale-aligned, shift-and-scale invariant, monocular, stereo, or multiview. ORB-SLAM likewise evaluates monocular trajectories after similarity alignment because absolute scale is not observable from one camera alone. Its bundle-adjusted feature map supports camera motion and sparse structure; it is not a dense metric scene model.

Neural scene representations

Section titled “Neural scene representations”
Representation Primary reading · year · venue/status Strength and caveat
neural radiance field NeRF — 2020 · ECCV, peer reviewed continuous density/color field with differentiable volume rendering; per-scene optimization and view-dependent ambiguity
3D Gaussian primitives 3D Gaussian Splatting — 2023 · ACM TOG/SIGGRAPH, peer reviewed explicit anisotropic Gaussians and fast rasterization; rendering quality need not imply accurate surfaces
pairwise point maps DUSt3R — 2024 · CVPR, peer reviewed learned point-map regression unifying depth, matching, pose, and reconstruction
feed-forward geometry VGGT — 2025 · CVPR, peer reviewed jointly predict cameras, depth, point maps, and tracks from one to many views
multimodal geometry OmniVGGT — 2026 · CVPR, peer reviewed optionally inject depth and camera cues; compare RGB-only and auxiliary-input evidence explicitly

Separate geometry from rendering. PSNR, SSIM, and LPIPS assess held-out appearance; pose error, depth metrics, point accuracy/completeness, Chamfer distance, or F-score assess geometry. Each metric also depends on alignment and valid-region rules.

Embodied perception

Section titled “Embodied perception”

An embodied system chooses or executes actions a_t based on observations o_≤t, language or goals g, and internal state. Perception, dynamics, planning, and control can fail separately:

π(a_t | o_≤t, g),     p(z_(t+1) | z_≤t, a_t)

For a vision-language-action or world-model paper, record:

  • observation modalities, camera viewpoints, frame history, and proprioception;
  • which language or goal information is available;
  • action representation and control frequency;
  • offline data, demonstrations, simulation, and robot embodiment;
  • open-loop prediction versus closed-loop success;
  • safety constraints, intervention rate, and recovery behavior.

A convincing generated future is not proof that a policy can act, and benchmark success on one embodiment is not general physical understanding.

Stage Primary reading · year · venue/status Scientific boundary
E1 PaLM-E — 2023 · ICML, peer reviewed injecting continuous sensor and vision features into an embodied language model; language transfer is not closed-loop control
E2 RT-2 — 2023 · CoRL, peer reviewed representing robot actions as tokens for vision–language–action transfer; embodiment, demonstrations, and control frequency bound the claim

Controlled labs

Section titled “Controlled labs”

Temporal lab

Section titled “Temporal lab”

Train or probe one video model under identical clips. Compare normal order, temporally shuffled frames, reversed order, single frames, and a scene-only baseline. Report accuracy by action type, temporal stride, number of clips/crops, compute, and latency.

Geometry lab

Section titled “Geometry lab”

Capture a small static scene with overlapping views:

  1. run a feature/SfM pipeline and a feed-forward learned geometry model;
  2. preserve original resolution and camera metadata;
  3. compare registered-view count, pose error where reference is available, reprojection error, depth/point completeness, and failure cases;
  4. render held-out views but report rendering separately from geometry;
  5. perturb calibration, image order, overlap, lighting, and texture.

Embodied audit

Section titled “Embodied audit”

For one public policy or world-model benchmark, map every observation to every action target. Run observation-drop, temporal-delay, and language-goal counterfactuals if artifacts permit. Report task success with confidence intervals, failure stage, horizon, interventions, and inference rate.

Failure modes

Section titled “Failure modes”
  • scene or object appearance solves an “action” benchmark without motion;
  • frames from one video or scene cross train and test splits;
  • test-time clips/crops differ substantially between compared models;
  • tracking quality is inherited from a stronger detector;
  • long-horizon predictions blur or drift while short-horizon metrics remain strong;
  • camera intrinsics, poses, or depth are supplied to only one 3D method;
  • scene-scale alignment hides metric-depth failure;
  • novel-view quality is presented as accurate geometry;
  • generated future plausibility replaces physical or causal evaluation;
  • a policy exploits simulator, embodiment, or language shortcuts;
  • offline and causal systems receive different future context.

Self-check and exit task

Section titled “Self-check and exit task”
  1. Which ablation reveals whether a video classifier uses temporal order?
  2. How do detection and association contribute separately to tracking?
  3. What is the distinction between pixel reconstruction and latent prediction?
  4. Which assumptions make triangulation possible?
  5. Why can excellent novel-view synthesis coexist with poor geometry?
  6. What additional test turns a predictive video model into evidence about control?

Exit task: produce one temporal-or-geometric evidence sheet. Include tensor and coordinate conventions, available modalities, causal context, training target, counterfactual tests, task metrics, efficiency, and three failure examples.

Pathology bridge

Section titled “Pathology bridge”

WSIs are not ordinary videos or perspective scenes, but temporal and 3D ideas still transfer carefully: serial sections, z-stacks, multiplex rounds, spatial molecular assays, multiresolution navigation, and multiple slides per case create structured sets. Preserve physical coordinates, registration uncertainty, specimen/case grouping, and acquisition order. Continue to WSI engineering, spatial and molecular modeling, and case-level MIL.

Calibrated metric pose and directional consensus under uncertainty

PRISM-VO combines calibrated plenoptic depth with temporal photometric residuals; LAYS separates yaw using column-to-line voting over uncertain positions. Compare metric observability, correspondence geometry and uncertainty handling instead of treating all pose estimation as one task.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle AdjustmentAymeric Fleith, Julian Zirbel, Daniel Cremers, and Niclas Zeller2026 · ECCV · acceptedAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. The official project announces forthcoming code; a runnable public release was not verified.PRISM-VO couples temporal photometric tracking with metric depth from a focused plenoptic camera. Its sparse sliding-window bundle adjustment jointly estimates camera poses, inverse point depths and brightness parameters, using uncertainty-weighted inverse virtual-depth residuals. The physical micro-lens geometry makes metric scale observable; it is not recovered from a learned monocular prior.Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record
Cross-View Yaw Estimation in Location Uncertainty with Line-Aligning Yaw ScoringTaeho Kang, Nairan Zhang, Yelin Kim, Yujiao Shi, and Youngki Lee2026 · ECCV · acceptedAccepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Exact KITTI measurements follow Table 2, which differs from the surrounding prose; that discrepancy is retained in the review.LAYS estimates camera yaw by matching ground-image columns to aerial-image pixels and accumulating line-consensus votes. Moving a candidate camera position along the same directed radial ray preserves the implied yaw, so orientation can remain identifiable while location is uncertain. The method still searches a 3D position–yaw tensor; it does not eliminate position candidates from the computation.Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record

Reading question: What anchors metric scale or orientation, and is translation estimated precisely or deliberately left ambiguous?

Evidence boundary: Approximate pinhole baselines and differing yaw/noise protocols require explicit qualifiers. Cross-view inference is not cross-validation or cross-resolution distillation; a measurement residual is not a residual teacher.

Physiology-aware cross-modal spectral distillation

RPM-Distill transfers peak position, off-peak suppression and spectral morphology from an RF teacher to an RGB student. A validation-driven policy chooses sample and loss weights; deployment removes the radar branch. Compare this privileged-sensor design with feature KD and reliability-weighted transfer.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
RPM-Distill: Physiology-guided Adaptive Cross-modal Distillation for Robust Remote Physiological MeasurementJiyao Wang, Qingyong Hu, Duoxun Tang, Xiao Yang, Kaishun Wu, and Jiangbo Yu2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10941. Publisher proceedings/version of record was not verified on 2026-09-12.RPM-Distill transfers rhythm information from a frozen radar teacher to a video-only physiological student. It compares band-limited spectra rather than forcing incompatible RF and RGB waveforms to match. A learned policy gates each sample and weights peak, off-peak and spectral-shape losses according to whether they improve held-out physiological recovery.ECCV Spotlight program · Conference manuscript · Official implementation

Reading question: Which teacher signal is physiologically invariant across sensors, and what labeled validation budget trains the transfer policy?

Evidence boundary: A supervised teacher and held-out labels are still required; the pipeline is not globally label-free. Policy residual blocks are not a residual-error KD target, and fusion methods retain different inference inputs.

Low-resolution student inference versus resolution-cascaded generation

GAP-Track removes teacher and residual reconstruction branches at inference. RMD instead retains high-resolution generative refinement. Both exploit cross-resolution supervision but solve different deployment problems.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
GAP-Track: Bridging the Resolution Gap for Cross-Resolution RGBT TrackingShiyu Zhang, Tianyang Xu, Zhangyong Tang, He Wang, Xiao-Jun Wu, and Josef Kittler2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. Table 1 governs the recreated LasHeR values; the quarter-resolution CKD SR differs between Table 1 (35.3) and Table 4 (35.5). The PGS angle/radial equations need origin and quadrant safeguards not specified in the displayed formulas. The repository contains implementation directories despite a remaining release-soon notice; code was inspected, not executed.GAP-Track transfers full-resolution tracking supervision to a low-resolution RGB–thermal student. A masked residual feature generator learns to approximate teacher features during training, and a hierarchy of feature teachers eases extreme downsampling. The deployed model keeps only the low-resolution student and box head; it does not first reconstruct a high-resolution image at inference.Official ECCV poster record · Official code

Reading question: Is high resolution absent at deployment, or only used for fewer steps? What is the full-resolution reference?

Evidence boundary: Synthetic downsampling does not reproduce all optical changes between microscope magnifications; lower-resolution gains need not equal full-resolution performance.

Volume tokens, persistent geometry and registration

Compare learned volume representations, streaming geometric context, flow-based rigid registration and dense image matching. Each chooses a different correspondence unit and computation path; learned priors complement rather than abolish geometric constraints.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Geometric Context Transformer for Streaming 3D ReconstructionLin-Zhuo Chen, Jian Gao, Shangzhan Zhang, Yihang Chen, Nan Xue, Jianyuan Wang, Christian Rupprecht, Xun Cao, Xing Zhu, Yujun Shen, Yao Yao, and Yinghao Xu2026 · ECCV · acceptedAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv/released system uses the LingBot-Map name and a different author list; authors and quantitative claims here follow the official ECCV manuscript.GCT is a causal 3D reconstruction Transformer whose memory is organized by geometric function: fixed anchors establish a reference frame, recent images supply dense local overlap, and compact trajectory tokens preserve historical context. This separates what must remain detailed from what can be compressed, rather than treating every past frame identically.Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code (LingBot-Map)

Reading question: Which coordinates, features or correspondences are predicted, and how are rigidity, memory growth and distribution shift tested?

Evidence boundary: Compact per-frame context still accumulates; multi-step integration is not one-pass inference, and improved average matching can hide benchmark regressions.

Predictive representations for reward-selected control

TD-JEPA connects non-contrastive latent prediction to temporal-difference learning. Read the state/action input contract and downstream task-vector inference before interpreting zero-shot control.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement LearningMarco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, Andrea Tirinzoni2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Reward-free state–action transitions, not action-free video. Zero-shot policy selection still requires reward observations for task-vector inference. Theoretical assumptions, benchmark-specific actor inputs and mixed OGBench comparisons are qualified in the source-checked review; no WSI evaluation.TD-JEPA learns state and task representations from reward-free transitions so that a new reward can select a policy without task-specific policy fine-tuning. Its distinguishing move is temporal-difference bootstrapping in a joint-embedding predictive architecture: predicting accumulated future task features connects representation learning to zero-shot control.Official proceedings record · OpenReview · Official implementation

Reading question: What future quantity is predicted, what prevents collapse, and what reward information is still needed at deployment?

Evidence boundary: Restricted theory and simulator-specific actor inputs do not prove arbitrary video or pathology-control validity.

Prompt-conditioned audiovisual embeddings

WAVE combines speech and sound pathways, layer-wise non-text fusion and joint retrieval/QA embedding training. Its output is similarity ranking, not unrestricted text generation.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
WAVE: Learning Unified & Versatile Audio-Visual Embeddings with Multimodal LLMChangli Tang, Qinfan Xiao, Ke Mei, Tianyi Wang, Fengyun Rao, Chao Zhang2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Multimodal embeddings and candidate-answer ranking, not unrestricted report generation. Official paper and released training configuration checked; unequal pooling contracts, negative findings and unresolved overall-score discrepancy are documented. No WSI evaluation.WAVE turns a multimodal generative backbone into a prompt-conditioned embedding model for text, video, audio and audiovisual inputs. Its central design combines complementary speech and sound encoders with hierarchical feature fusion across LLM layers. Retrieval and multiple-choice QA are both formulated as embedding comparisons, not unrestricted answer generation.Official proceedings record · OpenReview · Official implementation

Reading question: How do the task prompt and pooling rule change what information the embedding retains?

Evidence boundary: Modality coverage is finite; question-conditioned candidate accuracy is not evidence of report-generation reliability.

Diverse representatives and temporal token merging

FlashVID preserves representative video tokens and merges the unselected remainder into similarity trees. Compare token budgets, compression overhead and complete time to first token separately.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token MergingZiyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, Zhuotao Tian2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Inference-time token selection/merging, not residual-error KD. Full configuration includes segmentation, budget adjustment and hybrid LLM pruning. Report encoding and compression overhead as well as LLM savings; natural-video evidence is not cross-magnification pathology validation.FlashVID compresses video tokens without training the target Video LLM. Attention-guided Diverse Token Selection keeps representative tokens, while Tree-based Spatiotemporal Merging pools redundant tokens across adjacent frames. This is token selection and merging, not teacher–student knowledge distillation.Official proceedings record · OpenReview · Official implementation

Reading question: Which information survives selection or merging, and where is compression cost included?

Evidence boundary: Residual tokens are not residual-error KD; large LLM-only savings are not the same as end-to-end acceleration.

Complete maintained paper route

Section titled “Complete maintained paper route”

43 canonical readings · 2006–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered by an explicit prerequisite-aware teaching sequence. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.

Show all 43 readings in maintained order

  8. #82026ECCVAccepted · proceedings pendingL3

    RPM-Distill: Physiology-guided Adaptive Cross-modal Distillation for Robust Remote Physiological Measurement

    Jiyao Wang, Qingyong Hu, Duoxun Tang, Xiao Yang, Kaishun Wu, and Jiangbo Yu

    ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #10941. Publisher proceedings/version of record was not verified on 2026-09-12.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation

    Metadata checked

  11. #112026ICLRPeer reviewedL3

    TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning

    Marco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, Andrea Tirinzoni

    Official ICLR 2026 oral and full proceedings manuscript checked. Reward-free state–action transitions, not action-free video. Zero-shot policy selection still requires reward observations for task-vector inference. Theoretical assumptions, benchmark-specific actor inputs and mixed OGBench comparisons are qualified in the source-checked review; no WSI evaluation.

    Official artifacts: Official proceedings record · OpenReview · Official implementation

    Metadata checked

  20. #202026ECCVAccepted · proceedings pendingL3

    GAP-Track: Bridging the Resolution Gap for Cross-Resolution RGBT Tracking

    Shiyu Zhang, Tianyang Xu, Zhangyong Tang, He Wang, Xiao-Jun Wu, and Josef Kittler

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. Table 1 governs the recreated LasHeR values; the quarter-resolution CKD SR differs between Table 1 (35.3) and Table 4 (35.5). The PGS angle/radial equations need origin and quadrant safeguards not specified in the displayed formulas. The repository contains implementation directories despite a remaining release-soon notice; code was inspected, not executed.

    Official artifacts: Official ECCV poster record · Official code

    Metadata checked

  23. #232026ECCVAccepted · proceedings pendingL3

    PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle Adjustment

    Aymeric Fleith, Julian Zirbel, Daniel Cremers, and Niclas Zeller

    Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. The official project announces forthcoming code; a runnable public release was not verified.

    Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record

    Metadata checked

  24. #242026ECCVAccepted · proceedings pendingL3

    Cross-View Yaw Estimation in Location Uncertainty with Line-Aligning Yaw Scoring

    Taeho Kang, Nairan Zhang, Yelin Kim, Yujiao Shi, and Youngki Lee

    Accepted to ECCV 2026; oral presentation verified in the official program. The conference-hosted manuscript is the review authority; a separate Springer proceedings record was not verified. Exact KITTI measurements follow Table 2, which differs from the surrounding prose; that discrepancy is retained in the review.

    Official artifacts: Official ECCV oral presentation · Official ECCV paper record · Author project · arXiv record

    Metadata checked

  34. #342026ECCVAccepted · proceedings pendingL3

    Geometric Context Transformer for Streaming 3D Reconstruction

    Lin-Zhuo Chen, Jian Gao, Shangzhan Zhang, Yihang Chen, Nan Xue, Jianyuan Wang, Christian Rupprecht, Xun Cao, Xing Zhu, Yujun Shen, Yao Yao, and Yinghao Xu

    Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified. The arXiv/released system uses the LingBot-Map name and a different author list; authors and quantitative claims here follow the official ECCV manuscript.

    Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code (LingBot-Map)

    Metadata checked

  39. #392026ECCVAccepted · proceedings pendingL3

    Reconstructing Humans and Objects in Interaction using Large Reconstruction Models

    Agniv Chatterjee and Georgios Pavlakos

    ECCV 2026 poster presentation is now verified against the official conference inventory (poster 3826), superseding the earlier author-only acceptance evidence. A Springer version of record was not verified on 2026-09-12; the existing review remains tied to its declared manuscript revision.

    Official artifacts: Official project · Official implementation · Official ECCV 2026 poster

    Metadata checked

  40. #402026arXivPreprintL3

    PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

    Yuandong Pu, Le Zhuo, Sayak Paul, Gabriel Jorge Menezes, Avram Đorđević, Shiyang Li, Yifan Zhou, Bin Fu, Wenlong Zhang, Junjun He, Yu Qiao, Yihao Liu, Jingbo Xing, and Xi Chen

    arXiv v1 benchmark preprint submitted 27 August 2026; conclusions concern terminal-outcome distributions in its controlled scenarios.

    Official artifacts: Official project · Official implementation · Official dataset

    Metadata checked

  42. #422026ICLRPeer reviewedL3

    WAVE: Learning Unified & Versatile Audio-Visual Embeddings with Multimodal LLM

    Changli Tang, Qinfan Xiao, Ke Mei, Tianyi Wang, Fengyun Rao, Chao Zhang

    Official ICLR 2026 oral and full proceedings manuscript checked. Multimodal embeddings and candidate-answer ranking, not unrestricted report generation. Official paper and released training configuration checked; unequal pooling contracts, negative findings and unresolved overall-score discrepancy are documented. No WSI evaluation.

    Official artifacts: Official proceedings record · OpenReview · Official implementation

    Metadata checked

  43. #432026ICLRPeer reviewedL3

    FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging

    Ziyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, Zhuotao Tian

    Official ICLR 2026 oral and full proceedings manuscript checked. Inference-time token selection/merging, not residual-error KD. Full configuration includes segmentation, budget adjustment and hybrid LLM pruning. Report encoding and compression overhead as well as LLM savings; natural-video evidence is not cross-magnification pathology validation.

    Official artifacts: Official proceedings record · OpenReview · Official implementation

    Metadata checked