Skip to content
Computational Pathology Reading Roadmap

Vision evaluation and robustness

Evaluation is a measurement design problem. A model score is conditional on the data unit, split, annotation, preprocessing, adaptation, threshold, metric implementation, and hardware. Robust evidence makes these conditions explicit and tests plausible ways they can fail.

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.

Evaluation work is organized by validity threat—overlap, shift, calibration, subgroup performance, and efficiency—not by the date it was added.

Prerequisites and objectives

Section titled “Prerequisites and objectives”

Complete the evaluation and reproducibility material in ML foundations and at least one task chapter in this CV route. You should be able to:

  • define the independent sampling unit and prevent group leakage;
  • match a metric to the intended visual decision;
  • quantify uncertainty rather than compare unsupported decimal places;
  • distinguish corruption, domain, subpopulation, temporal, and label shift;
  • evaluate calibration and selective prediction;
  • measure latency, throughput, memory, and energy without hiding preprocessing;
  • produce a model and dataset record another group can audit.

Begin with the estimand

Section titled “Begin with the estimand”

Write the target population, prediction unit, outcome, operating point, and loss before choosing a benchmark. Examples:

  • expected top-1 error for photographs from named sources;
  • lesion sensitivity at a fixed false-positive rate per image;
  • instance recall for small objects;
  • mean boundary error under a specified annotation protocol;
  • selective risk when the model may abstain.

If images i belong to subjects, videos, scenes, websites, or acquisition sessions g(i), splitting images independently estimates the wrong generalization problem. Group first, then split.

Split and provenance audit

Section titled “Split and provenance audit”

For every dataset, record:

  1. source and collection process;
  2. license, consent, privacy, and permitted redistribution;
  3. image, group, class, and subgroup counts;
  4. annotation instructions, annotators, disagreement, and missingness;
  5. duplicate and near-duplicate policy;
  6. train/validation/test construction and all tuning access;
  7. preprocessing, exclusions, label mappings, and version;
  8. possible pretraining or benchmark overlap.

Model cards and dataset documentation are experimental controls, not administrative appendices. Use Model Cards and Datasheets for Datasets as reporting anchors.

Task-to-metric map

Section titled “Task-to-metric map”
Task Primary metrics Required complementary evidence
classification accuracy, balanced accuracy, macro F1, AUROC/AUPRC when suitable confusion matrix, class frequency, calibration, operating threshold, subgroup slices
detection COCO-style AP across IoU thresholds AP50/AP75, small/medium/large AP, recall, localization/classification/duplicate errors
semantic segmentation per-class and mean IoU, Dice boundary metric, rare-class recall, empty-mask convention, object-level errors
instance/panoptic segmentation mask AP or panoptic quality recognition versus segmentation quality, size/count slices, merge/split errors
retrieval Recall@K, median rank, mAP query strata, candidate-pool size, duplicate policy, confidence or abstention
depth / optical flow AbsRel/RMSE/threshold accuracy or EPE/outlier rate scale/alignment protocol, valid pixels, motion/depth ranges, boundary performance
tracking HOTA, IDF1, MOTA where historically needed detector AP, identity switches, fragmentation, occlusion and duration slices
3D reconstruction pose, depth, point/surface accuracy and completeness alignment, registered-view rate, failure rate; rendering metrics reported separately
generation likelihood/ELBO where available, FID/KID, precision/recall sample count, feature extractor, memorization, conditional compliance, blinded human study, sampling cost

Average metrics require an aggregation definition. Macro over classes, micro over pixels, mean over images, and pooled counts answer different questions. For FID, record the exact Inception implementation, resize and normalization path, reference split, generated sample count, and random seed. The original definition compares Gaussian fits in one feature space; it is neither a per-image score nor a complete test of diversity, memorization, or semantic correctness.

Calibration, uncertainty, and abstention

Section titled “Calibration, uncertainty, and abstention”

For confidence c_i and correctness a_i, expected calibration error bins predictions and estimates

ECE = Σ_(b=1)^B [ |I_b| / n · |acc(I_b) − conf(I_b)| ]

ECE is bin-dependent and can hide class or subgroup miscalibration. Include proper scores such as negative log-likelihood or Brier score, reliability diagrams, and task-specific calibration.

Primary reading · year · venue/status Read it for
On Calibration of Modern Neural Networks — 2017 · ICML, peer reviewed reliability diagrams, ECE, and post-hoc temperature scaling
Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles — 2017 · NeurIPS, peer reviewed predictive diversity and uncertainty from independently trained models
A Baseline for Detecting Misclassified and Out-of-Distribution Examples — 2017 · ICLR, peer reviewed maximum-softmax confidence as a minimum OOD control, not a deployment guarantee

For abstention, plot coverage against risk. A model that is accurate only after rejecting difficult subgroups may be unsafe even when overall selective risk looks good.

Corruption and distribution shift

Section titled “Corruption and distribution shift”

Corruption benchmarks transform the same underlying images; resampled test sets expose benchmark sampling uncertainty; cross-domain tests change acquisition, context, or population.

Stage Primary reading · year · venue/status Validity threat
R1 Gender Shades — 2018 · FAT, peer reviewed* intersectional subgroup disparity and unbalanced commercial evaluation
R2 Benchmarking Neural Network Robustness to Common Corruptions and Perturbations — 2019 · ICLR, peer reviewed severity-controlled common corruptions and perturbation consistency
R3 Do ImageNet Classifiers Generalize to ImageNet? — 2019 · ICML, peer reviewed performance change on a new test set constructed for the same task
R4 ObjectNet — 2019 · NeurIPS, peer reviewed a test-only collection that randomizes object background, rotation, and viewpoint; six ImageNet-trained models lost roughly 40–45 percentage points on overlapping classes
R5 ImageNet-trained CNNs are Biased towards Texture — 2019 · ICLR, peer reviewed shape–texture conflict and shortcut-sensitive evaluation
R6 WILDS — 2021 · ICML, peer reviewed realistic distribution shifts with group and domain metadata

ObjectNet is not an ordinary retraining benchmark: the paper intentionally supplies no paired training set. Treating its controlled test images as development data changes the estimand from out-of-distribution generalization to adaptation on a known target collection.

Do not tune on the external test and continue calling it external. Once examples, metrics, or failure cases guide development, create a new untouched evaluation or describe the set as a development domain.

Open-set and adversarial robustness

Section titled “Open-set and adversarial robustness”

OOD detection, corruption robustness, and adversarial robustness test different threats. An OOD score needs named in/out populations and an operational threshold. An adversarial evaluation needs a threat model: norm or transformation set, radius, attacker knowledge, iterations, restarts, targeted/untargeted goal, and whether preprocessing and stochasticity are inside the attack.

Stage Primary reading · year · venue/status Scientific boundary
A1 Explaining and Harnessing Adversarial Examples — 2015 · ICLR, peer reviewed the fast gradient sign method, local-linearity hypothesis, and an attack precursor that must be separated from a defense claim
A2 A Baseline for Detecting Misclassified and Out-of-Distribution Examples — 2017 · ICLR, peer reviewed maximum softmax probability is a reproducible baseline but can remain confident far from the training distribution
A3 Towards Deep Learning Models Resistant to Adversarial Attacks — 2018 · ICLR, peer reviewed projected-gradient adversarial training under a declared first-order threat model
A4 The Geometry of Robustness — 2026 · CVPR, peer reviewed joint ID, OOD, and adversarial fine-tuning claims for VLMs; reproduce all three axes and matched adaptation budgets

Robust losses are not universal defenses

Section titled “Robust losses are not universal defenses”
Primary reading · year · venue/status Assumption to test
Symmetric Cross Entropy for Robust Learning with Noisy Labels — 2019 · ICCV, peer reviewed whether the bidirectional entropy construction helps under realistic instance- or class-dependent label noise, not only synthetic flips
A General and Adaptive Robust Loss Function — 2019 · CVPR, peer reviewed whether learning the residual-distribution shape improves clean and corrupted data without hiding scale or optimizer changes

Attack success against an undefended model is not evidence that a proposed defense is robust. Re-run adaptive attacks, include clean accuracy, check gradient masking, and separate security claims from ordinary distribution-shift generalization.

Counterfactual and shortcut tests

Section titled “Counterfactual and shortcut tests”

A useful audit changes one hypothesized shortcut while preserving intended evidence:

  • mask or replace backgrounds;
  • stylize texture while preserving shape;
  • remove text, metadata, borders, timestamps, or watermarks;
  • shuffle frames or reverse temporal order;
  • perturb camera calibration or view overlap;
  • exchange prompts, synonyms, attributes, or class order;
  • crop to the claimed region and, separately, remove it;
  • equalize source, resolution, or class-frequency cues.

Counterfactuals can also introduce artifacts, so include positive and negative controls and inspect images rather than trusting an automated transform.

Statistical comparison

Section titled “Statistical comparison”

Report uncertainty at the independent data unit. For grouped images, resample groups rather than pixels or tiles. Paired resampling is preferable when two systems are tested on the same examples.

Pre-specify:

  • primary metric and direction;
  • aggregation and missing-output rule;
  • confidence-interval or resampling method;
  • number of seeds for stochastic training;
  • subgroup and multiple-comparison policy;
  • minimum practically important difference.

Variation across random seeds and uncertainty from a finite test population are different. Report both when training is repeated.

Efficiency is empirical

Section titled “Efficiency is empirical”

Parameters and FLOPs omit activations, memory traffic, unsupported kernels, dynamic shapes, preprocessing, host/device transfer, decoding, and postprocessing.

Measure:

  • batch-one latency and a stated throughput batch;
  • warm-up, repetitions, synchronization, and variability;
  • end-to-end and model-only time;
  • peak accelerator and host memory;
  • input resolution, clip/view count, token or proposal count;
  • hardware, software versions, precision, compilation, and power mode;
  • preprocessing, retrieval, decoding, NMS, sampling steps, or 3D optimization;
  • training or distillation cost when making lifecycle claims.

Use the intended operating point. A detector timed with few outputs, a VLM without image preprocessing, or a diffusion model without its autoencoder is incomplete.

Controlled robustness lab

Section titled “Controlled robustness lab”

Choose one trained model from another chapter and freeze its checkpoint.

  1. Reproduce the internal test metric from a saved prediction file.
  2. Deduplicate and verify the group-level split.
  3. Evaluate at least three corruptions over severity, one new source/domain, and meaningful class/subgroup slices.
  4. Plot reliability and compute a proper score before and after temperature scaling fitted only on validation data.
  5. Run one counterfactual designed around an observed shortcut.
  6. Bootstrap the independent unit for paired confidence intervals.
  7. measure end-to-end latency, throughput, and memory on named hardware;
  8. create a failure gallery sampled by a pre-declared rule, not cherry-picking.

The final report should separate descriptive findings from causal explanations. An error correlation suggests a mechanism; an intervention may test it.

Failure modes

Section titled “Failure modes”
  • random image splits leak subjects, scenes, videos, or near-duplicates;
  • model or prompt selection uses the test set;
  • overall accuracy hides rare-class or subgroup collapse;
  • AUROC is used despite an operating point dominated by severe class imbalance;
  • empty masks or missing predictions receive inconsistent scores;
  • ECE alone is presented as complete calibration;
  • an external domain becomes a repeatedly tuned development set;
  • confidence intervals resample pixels or tiles rather than independent groups;
  • one model receives test-time augmentation, more crops, higher resolution, or a larger retrieval corpus;
  • latency omits preprocessing or asynchronous synchronization;
  • a benchmark’s web images overlap foundation-model pretraining;
  • robustness on synthetic corruption is generalized to all real shifts.

Self-check and exit task

Section titled “Self-check and exit task”
  1. What is the independent sampling unit for your current dataset?
  2. Which metric directly represents the intended decision and operating point?
  3. What does calibration add beyond ranking or accuracy?
  4. How do corruption, resampling, and domain shift differ?
  5. Which counterfactual would falsify your leading shortcut explanation?
  6. What parts of the deployed pipeline are absent from model-only FLOPs?

Exit task: produce an auditable evaluation packet containing dataset and split records, immutable predictions, metric code/version, confidence intervals, calibration, subgroup and shift tables, counterfactual evidence, failure gallery, and an end-to-end efficiency record.

Pathology bridge

Section titled “Pathology bridge”

Pathology magnifies every validity threat: tiles share patients and slides; scanner, site, stain, organ, and time create shifts; labels may be report-derived; rare findings drive clinical utility; and pretraining cohorts can overlap public benchmarks. Apply patient-level resampling and provenance, then continue to pathology evaluation and clinical translation.

Physics-informed optical fields and focus-stack geometry

Event-domain refractive inversion estimates evolving air density, while calibrated focus-stack attention estimates scene depth. Both exploit a known optical observation process; compare acquisition metadata, scene-specific optimization and synthetic-to-real validation before treating them as generic image encoders.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Zero-shot Depth from DefocusYiming Zuo, Hongyu Wen, Venkat Subramanian, Patrick Chen, Karhan Kayan, Mario Bijelic, Felix Heide, and Jia Deng2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12.FOSSA learns metric depth from a calibrated focus stack using attention across focus settings. It alternates ordinary image-token processing with per-location stack attention conditioned on focus distance, then pools the stack and predicts dense depth. The accompanying ZEDD benchmark tests real-world zero-shot transfer rather than within-dataset fitting.ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)

Reading question: Which unknown is identifiable from the actual measurements, and what boundary, calibration or scene-static assumptions supply missing information?

Evidence boundary: High event acquisition rate is not real-time optimization; multiple focus images provide different evidence from one monocular image.

Resolution transfer and calibrated geophysical fields

BaguanHR creates additional fine-resolution atmospheric training fields through super-resolution, whereas Sparse2Cloud predicts dense cloud geometry from sparse active-sensor tracks and calibrates quantile intervals. Compare transferred information, geographic or temporal splits and uncertainty, not just grid density.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Ice Cloud Geometry Retrieval with Calibrated Uncertainty from Passive Satellite ImageryAyush Prasad2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #12985. Publisher proceedings/version of record was not verified on 2026-09-12. Selected aggregate values follow Table 1. Table 2 reports a different G-Q aggregate R2 (0.725 vs 0.731); keep table-specific provenance. CQR guarantees marginal coverage only under exchangeability, not arbitrary subgroups or global domain shift.Sparse2Cloud learns dense ice-cloud geometry from narrow satellite supervision tracks and attaches conformal prediction intervals. A ConvNeXt U-Net predicts eight cloud targets and three quantiles per target, conditioned on passive VIIRS channels and ERA5 profiles. A five-model quantile ensemble plus calibration separates point accuracy from interval coverage.ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)

Reading question: Does a finer output represent measured extra information, learned prior interpolation or a calibrated uncertainty statement?

Evidence boundary: Synthetic resolution does not guarantee missing fine physics. Marginal conformal coverage assumes exchangeability and is not a universal conditional guarantee under climate or geographic shift.

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.

Counterexamples for spatial, semantic and action grounding

Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
SOCO: Benchmarking Semantic Object Correspondence in Vision Foundation ModelsOlaf Dünkel, Basavaraj Sunagad, Haoran Wang, David T. Hoffmann, Christian Theobalt, and Adam Kortylewski2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #349 checked on 2026-09-12. Publisher version of record was not independently verified.SOCO benchmarks whether a representation matches the correct object part, not merely the correct kind of part. Its taxonomy separates concept correspondence, identity-specific semantic object correspondence and cross-category transfer. Language descriptions also let VLMs be tested on text-guided localization versus visual-reference matching.ECCV Spotlight program · Conference manuscript · Official project or implementation
Do Not Leave a Gap: Hallucination-Free Object Concealment in Vision-Language ModelsAmira Guesmi and Muhammad Shafique2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #12081 checked on 2026-09-12. Publisher version of record was not independently verified.Background-Consistent Re-encoding (BCR) studies how a vision-language model can stop naming a selected object without filling its missing representation with an invented replacement. It perturbs the input image so object-region features resemble surrounding background features while retaining the encoder's ordinary token layout. The scientific claim concerns continuity of visual evidence, not literal elimination of every hallucination.ECCV Spotlight program · Conference manuscript

Reading question: Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?

Evidence boundary: A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations.

Recovering, steering and testing visual concepts

Intermediate features can expose information hidden by the final representation. Compare frozen-backbone negation adaptation, language-directed representations, circuit interventions and geometry-subspace probes; distinguish information accessibility from causal control.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
“World Knowledge” in the Weights: Reading Concept Circuits of Vision TransformersYanlin Chen, Tang Li, and Xi Peng2026 · ECCV · acceptedECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #8168. Publisher proceedings/version of record was not verified on 2026-09-12.VisionCLT adapts cross-layer transcoders to visualize how sparse concepts interact inside a ViT. It offers a global circuit assembled from learned encoder/decoder connections and an instance circuit based on active concepts for an image. These are complementary approximations to the model's computation, useful for testing shortcuts and interventions rather than merely highlighting salient pixels.ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript

Reading question: Which layer, token, subspace or circuit changes, and does the intervention preserve unrelated capabilities?

Evidence boundary: A probe may measure accessibility rather than semantic understanding; intervention and natural distribution shift are different tests.

Description-mediated transfer for image quality

RALI transfers teacher quality descriptions into a CLIP encoder, then fits compressed score bases. Read its image-only deployment path separately from the paper’s multi-domain RACT procedure.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality AssessmentShijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li Zhang, Tianfan Xue, Jian Zhang2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Teacher-generated quality descriptions align a compact image encoder; deployed scoring uses PCA and fitted bases without an LLM. Keep RALI separate from multi-domain RACT. Dataset-specific trade-offs, causal-interpretation limits and the conflicting memory-reduction claim are qualified. No WSI or cross-magnification evaluation.RALI transfers quality-specific descriptions from a reasoning-based image quality assessment teacher into a CLIP image encoder. The deployed scorer does not generate a reasoning trace or load the teacher LLM: it compares a compressed image embedding with learned quality bases and combines their scalar scores.Official proceedings record · OpenReview · Official implementation

Reading question: Which teacher information is transferred, which modules disappear at inference, and which domains lose performance?

Evidence boundary: Comparable aggregate IQA performance is not universal replacement or WSI validation; account for offline description generation and the measured memory ratio.

Uncertainty contracts: confidence, marginal coverage and set efficiency

Temperature scaling preserves class ranking but changes adaptive conformal scores. A temperature that improves confidence calibration can enlarge prediction sets while reducing empirical class-coverage disparity. Use separate objectives and report all three axes instead of calling a model simply calibrated.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
On Temperature Scaling and Conformal Prediction of Deep ClassifiersLahav Dabah and Tom Tirer2025 · ICML · peer-reviewedCalibrated confidence and conformal prediction sets solve different uncertainty problems. Temperature scaling may make confidence more reliable yet enlarge adaptive prediction sets. Dabah and Tirer show this trade-off empirically and explain a non-monotonic temperature effect; they propose separate temperatures for confidence and prediction-set goals.Final proceedings PDF · Author implementation · Author manuscript

Reading question: Is the target honest confidence, a small candidate set, or reliable coverage for a specified population?

Evidence boundary: Marginal exchangeability-based coverage is not per-class, per-hospital or per-patient certainty. Temperature selection, calibration units and final test data must remain distinct.

Global semantic anchors and local patch targets

Two Global Crops Suffice separates same-image global-view alignment from local crops and iBOT refinement in controlled DINO-family experiments. CODI instead changes the patch target: weakly mix the same teacher view’s class distribution into its local patch distribution. The diagnostic study and target-side method answer related questions, but neither independently validates the other.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Two Global Crops Suffice: Locating Semantic Emergence in DINO-Style Self-Supervised LearningBasavaraj Sunagad, Artur Jesslen, and Adam Kortylewski2026 · arXiv · preprintarXiv:2609.28187v1, submitted 23 September 2026, inspected 26 October 2026. Empirical DINO-family ablations on ImageNet-1k, not a universal necessity theorem or an evaluation of DINOv3/pathology. The review retains a discrepancy between the main text’s fixed-scale depth claim and Appendix Table A3.This is a controlled diagnosis of DINO-style learning, not a new foundation model. In the tested ImageNet-1k regime, matching two geometrically different global crops from the same image already produces semantic patch correspondence. iBOT’s masked-patch objective improves that structure when combined with the global objective, while extra local crops chiefly improve classification in the DINO-only comparison. The practical lesson is to separate semantic anchoring, refinement and evaluation rather than treat a successful recipe as an indivisible block.Inspected manuscript and appendices (v1)

Reading question: Which signal creates useful spatial semantics, which refines it, and do your global and dense evaluations agree?

Evidence boundary: Both inspected sources are preprints. Crop ablations match iterations, not all computation; the depth-scale narrative has an appendix exception. CODI’s gradient statement is local/logit-level, its alpha is task-dependent, and its code was not released at inspection. Neither reports WSI or cross-magnification validation.

Complete maintained paper route

Section titled “Complete maintained paper route”

39 canonical readings · 2015–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. 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 39 readings in maintained order

  18. #182021ICLRPeer reviewedL2–L3

    Long-Tail Learning via Logit Adjustment

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar

    Metadata checked

  21. #212021IEEE Transactions on Image ProcessingPeer reviewedL3

    Resolution-Aware Knowledge Distillation for Efficient Inference

    Zhanxiang Feng, Jianhuang Lai, and Xiaohua Xie

    Peer-reviewed journal article, volume 30, pages 6985–6996. The paper calls its framework RKD, meaning resolution-aware—not relational—knowledge distillation. Teacher and student use the same architecture but paired high- and low-resolution inputs; only the low-resolution student is needed at inference.

    Official artifacts: PubMed record

    Metadata checked

  24. #242021Communications of the ACMPeer reviewedL1–L2

    Datasheets for Datasets

    Timnit Gebru et al.

    Metadata checked

  27. #272026Findings of EMNLPAccepted · proceedings pendingL3

    MVC-Bench: Benchmarking Calibration of Medical Vision-Language Models

    Ashshak Sharifdeen, Shihab Aaqil Ahamed, Ufaq Khan, Muhammad Akhtar Munir Sujair Ibrahim, Mohamed Rafeek Mareer Ahamed, Yutong Xie, Imran Razzak, and Muhammad Haris Khan

    The arXiv record reports acceptance to Findings of EMNLP 2026. No final ACL Anthology record was available when checked on 30 August 2026, so claims were verified against arXiv v1 and the official repository.

    Official artifacts: Official implementation

    Metadata checked

  29. #292026arXivPreprintL3

    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

  33. #332026ECCVAccepted · proceedings pendingL3

    Zero-shot Depth from Defocus

    Yiming Zuo, Hongyu Wen, Venkat Subramanian, Patrick Chen, Karhan Kayan, Mario Bijelic, Felix Heide, and Jia Deng

    ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #3173. Publisher proceedings/version of record was not verified on 2026-09-12.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)

    Metadata checked

  34. #342026ECCVAccepted · proceedings pendingL3

    Ice Cloud Geometry Retrieval with Calibrated Uncertainty from Passive Satellite Imagery

    Ayush Prasad

    ECCV 2026 Spotlight verified in the official program, session Computational, Medical & Scientific Imaging; scientific claims use conference-hosted manuscript #12985. Publisher proceedings/version of record was not verified on 2026-09-12. Selected aggregate values follow Table 1. Table 2 reports a different G-Q aggregate R2 (0.725 vs 0.731); keep table-specific provenance. CQR guarantees marginal coverage only under exchangeability, not arbitrary subgroups or global domain shift.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · Author project (paper/program link)

    Metadata checked

  35. #352026ECCVAccepted · 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

  36. #362026ECCVAccepted · proceedings pendingL3

    SOCO: Benchmarking Semantic Object Correspondence in Vision Foundation Models

    Olaf Dünkel, Basavaraj Sunagad, Haoran Wang, David T. Hoffmann, Christian Theobalt, and Adam Kortylewski

    ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #349 checked on 2026-09-12. Publisher version of record was not independently verified.

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

    Metadata checked

  38. #382026ICLRPeer reviewedL3

    Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment

    Shijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li Zhang, Tianfan Xue, Jian Zhang

    Official ICLR 2026 oral and full proceedings manuscript checked. Teacher-generated quality descriptions align a compact image encoder; deployed scoring uses PCA and fitted bases without an LLM. Keep RALI separate from multi-domain RACT. Dataset-specific trade-offs, causal-interpretation limits and the conflicting memory-reduction claim are qualified. No WSI or cross-magnification evaluation.

    Official artifacts: Official proceedings record · OpenReview · Official implementation

    Metadata checked

  39. #392026arXivPreprintL3

    Two Global Crops Suffice: Locating Semantic Emergence in DINO-Style Self-Supervised Learning

    Basavaraj Sunagad, Artur Jesslen, and Adam Kortylewski

    arXiv:2609.28187v1, submitted 23 September 2026, inspected 26 October 2026. Empirical DINO-family ablations on ImageNet-1k, not a universal necessity theorem or an evaluation of DINOv3/pathology. The review retains a discrepancy between the main text’s fixed-scale depth claim and Appendix Table A3.

    Official artifacts: Inspected manuscript and appendices (v1)

    Metadata checked