Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Automatic Discovery and Assessment of Interpretable Systematic Errors in Semantic Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that a purely inference-driven pipeline can automatically find human-interpretable systematic precision errors in semantic segmentation models on unlabelled data, with up to 95.23% accuracy on the person class.

desk verdict Useful proof-of-concept for slice discovery in segmentation, but the headline numbers measure human agreement with the pipeline's own criteria, not independently validated systematic errors. read the letter →

arxiv 2411.10845 v1 pith:CXDP6OBW submitted 2024-11-16 cs.CV

Jaisidh Singh , Sonam Singh , Amit Arvind Kale , Harsh K Gandhi This is my paper · ORCID
classification cs.CV
keywords semanticsegmentationsystematicerrorsslicediscoveryopen-vocabularydetectionCLIPBLIP-2precisionautonomousdriving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that slice discovery—finding coherent sub-groups on which a model fails systematically—can be done for semantic segmentation without any labelled test data. It proposes a training-free pipeline that first crops every region a segmentation model labels as a vulnerable road user, uses an open-vocabulary detector to flag patches that do not contain that concept, and then uses CLIP and BLIP-2 to check whether the flagged errors share a human-interpretable concept with a small neighbourhood of similar patches. The reported success rate for recovering such systematic errors reaches 95.23% for the person class on the in-distribution benchmark, and the same pipeline rejects non-systematic errors on an out-of-distribution adverse-weather dataset. If this holds, model auditing for autonomous driving no longer needs expert pixel-level annotation to find a model's repeating mistakes.

What carries the argument

The load-bearing object is the scoring function $\Omega(p)=1$ when $\sigma_1+\sigma_2-\sigma_3 \geq \alpha$, where $\sigma_1$ is the average CLIP cosine similarity between the query patch's text embedding and its $q$ nearest-neighbour image embeddings, $\sigma_2$ is the average sentence-encoder similarity between the query caption and the neighbour captions, and $\sigma_3$ is the similarity between the query caption and a prompt for the target class. This single threshold combines two consistency signals—the error must look like its neighbours and be described like them—and one exclusion signal—it must not actually be the target concept. The initial candidate set is produced by an open-vocabulary detector that declares a precision error whenever it returns no bounding box for the target class, and the threshold $\alpha$ is set empirically to 0.35 with neighbourhood size $q=3$.

What would settle it

An experiment that would settle it: replace the open-vocabulary detector in the pipeline with an independent detector whose failure modes are known to differ, and re-run the systematicity check; if the discovered error groups change substantially, the errors are detector artefacts, while if they persist, the signal comes from the segmentation model. A complementary check is to pixel-annotate a sample of the discovered groups; if most 'systematic error' patches actually contain the target class at the pixel level, the claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that systematic precision errors of a semantic segmentation model can be discovered zero-shot, from raw images alone, by composing a detector's 'no box' response with a neighbourhood-consistency check. A patch is a precision error if a pretrained open-vocabulary detector finds no instance of the target class inside it; the patch is a systematic error if it is conceptually linked to its nearest neighbours in both CLIP image-text space and BLIP-2 caption space, and its captions do not match the target class. The paper reports this pipeline identifies interpretable systematic error groups—snow and car regions mistaken for 'person', metal and car parts mistaken for 'bicycle'—with human agreement ranging from roughly 57% to 95% depending on model, detector, and dataset. The framework is restricted to precision errors for the 'person' and 'bicycle' classes, and it is designed to work on any semantic segmentation model without retraining.

Load-bearing premise

The decisive premise is that when a pretrained object finder finds no object of the target class inside a patch, the patch really is a mistake by the segmentation model; the paper's own numbers show that premise holds only about half to three-quarters of the time.

Editorial extensions

If this is right

  • If the central claim holds, slice discovery no longer requires labelled test data or training a new model, so the same audit tool can be pointed at any deployed segmentation model on raw fleet data.
  • The human-interpretable captions produced for each systematic error group give a concrete target for intervention, such as adding snowy-person or car-with-person patches to training sets.
  • Because the framework is modular, swapping the detector or the captioner should let the audit run on new semantic classes and new domains without retraining.
  • The reported out-of-distribution results imply that even when a model's errors are mostly random, the pipeline can reject them, which is the behaviour needed to avoid false alarms in deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A consequence the paper leaves implicit is that the discovered 'systematic errors' are only as trustworthy as the open-vocabulary detector: with detector accuracy between 49.45% and 75.1% against IoU ground truth, a quarter to half of the patches entering the systematicity check may be mislabelled, and the final groups could partly reflect the detector's regularities rather than the segmentation m
  • The neighbourhood-consistency criterion may over-reward the captioner's own stereotypes: if BLIP-2 describes both a snowy person and a snowy pole as 'snow', the method will group them as conceptually linked even though the underlying failure causes differ, so the discovered slices may be concept-coherent without being cause-coherent.
  • A testable extension would be to run the same pipeline with an independent detector whose failure modes are known to differ; if the discovered systematic error groups stay the same, the signal comes from the segmentation model, whereas if they change, the pipeline is auditing the detector.
  • The paper's reliance on human agreement from a small evaluator pool (two or three annotators) suggests a natural follow-on experiment: a larger-scale study with expert annotators and inter-annotator agreement metrics would clarify how much of the reported accuracy is driven by the pipeline versus by evaluator subjectivity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a training-free, modular pipeline that discovers human-interpretable systematic precision errors in semantic segmentation models on unlabelled data. Given a target class c_j (person, bicycle), it extracts SSM-predicted patches, uses an open-vocabulary detector (GroundingDINO or Owl-ViT) to flag patches that do not contain c_j as precision errors, and then scores each candidate patch by combining CLIP-based image-text similarity to its nearest neighbours (σ1), sentence-encoder caption similarity (σ2), and a negative penalty σ3 for the patch actually depicting c_j. Patches with σ1+σ2−σ3 ≥ α are labelled interpretable systematic errors. Experiments on BDD (in-distribution) and ACDC (out-of-distribution) with UperNet-ConvNeXt and UperNet-Swin report detector precision-error identification accuracy of 49.45–75.1% and human-evaluated systematic-error assessment accuracy up to 95.23%, plus ablations over patch size a and neighbourhood size q.

Significance. If the central claim holds, the paper would provide a useful, label-free tool for surfacing semantically coherent failure modes in segmentation models, an area that lags behind image-classification slice discovery. The strengths are real: the method is training-free and modular, uses only foundation-model inference, and the authors include ablations over two detectors, two patch sizes, and three neighbourhood sizes, with qualitative examples showing coherent error groups (e.g., snow and car parts misclassified as person). However, the validation strategy is the weak link: the headline accuracies measure agreement between the pipeline and a small, partially circular human study, not whether the discovered slices are actually systematic errors of the segmentation model in the sense defined in Section 2.

major comments (3)
  1. [Section 4.4, Table 2] The central quantitative claim—that the framework discovers systematic errors with up to 95.23% accuracy—is an agreement rate between the algorithm and a human study whose criteria mirror the pipeline's own criteria (patch does not depict c_j, neighbours are coherent, caption fits). It is not a demonstration that the SSM has significantly lower performance on these slices. Section 2 defines systematic errors as slices on which model performance is significantly lower, but the paper never computes the SSM's error rate on the discovered positive set, never compares against a class-average or random-slice baseline, and never validates the final positives against the IoU ground-truth labels that are already available and used in Table 1. Without such a comparison, the term 'systematic error' is not independently established; I strongly recommend adding an oracle or proxy validation: e.g., measure pixel-level precision of the SSM on the discovered slices versus matched random patches.
  2. [Section 3.2, Table 1] The error set entering the systematicity stage is substantially noisy. Table 1 reports precision-error identification accuracy of 49.45–75.1%, with the lowest value for ACDC bicycle. Under this accuracy, a large fraction of patches labelled as errors are actually correct detections missed by the detector (or vice versa). Criterion 2 (low σ3) provides some filtering, but the paper does not quantify how much detector noise remains in the final positive set. This matters because discovered 'systematic errors' could reflect the detector's or CLIP's regularities rather than the SSM's. Please report the accuracy of the final Ω(p)=1 set against the IoU ground truth (which is already computed for Table 1), or otherwise show that the detected slices are robust to detector choice and threshold. The current ACDC results, where the authors assert that no systematic errors exist, do not mitigate this because high accuracy there is dominated by true negatives.
  3. [Section 4.4 and Supplementary (Systematic error evaluation)] The human evaluation is internally inconsistent and too fragile to support the headline numbers. The main text states that 'we use 3 evaluators such that an annotation is made only when 2 or more evaluators agree,' while the supplementary section 'Systematic error evaluation' states that '2 human evaluators with normal/correct vision were used.' No inter-annotator agreement is reported, no confidence intervals are given, and the sample sizes per cell of Table 2 are not stated. Furthermore, the human raters are asked to verify essentially the same three conditions that the pipeline optimizes, so the evaluation is partly circular. Please clarify the exact number of evaluators, report per-class sample sizes and agreement statistics, and add a validation that does not share the pipeline's decision rule—for example, asking raters to judge whether the SSM actually erred on the patch against ground-truth labels.
minor comments (5)
  1. [Table 2 and Supplementary Tables] There is a typographical inconsistency in the column header of Table 2: 'ConNeXt' should be 'ConvNeXt' to match the main text and the rest of the tables.
  2. [Section 3.3.2, Eq. (1)] The definition of σ1 computes cosine similarity between the CLIP text embedding of the query caption T_p and the CLIP image embeddings of the neighbours, but the notation E'_p = h_text(T_p) is used only once; please define it locally and ensure the subscript/superscript conventions for E_p and E_N are consistent, since E_p is reused for the sentence-encoder embedding in Eq. (2).
  3. [Section 4.3, Eq. (4)] The threshold α=0.35 is described as 'ascertained empirically: by observing the ranges of cosine similarities in each latent space used.' This is not a reproducible criterion; please provide the observed ranges, the number of samples inspected, and a sensitivity analysis around α, especially since the score is a sum of three cosine similarities that may have different scales.
  4. [Supplementary, Ablation Studies] Several figure references are broken or missing in the supplementary (e.g., 'Fig. ??' for the bicycle a=40 case), and Tables 10 and 11 are both numbered 10 while the text refers to 'Table 10' twice; please fix the numbering and the cross-references.
  5. [Section 4.4] The sentence 'In ACDC, patches do not contain systematic errors (verified by our human evaluation)' is presented as established fact, but the human study evaluates algorithmic predictions, not the ground-truth existence of systematic errors; please soften this claim or provide the supporting evidence.

Circularity Check

2 steps flagged · score 6.0 of 10

Table 2's headline accuracy is a human-algorithm agreement on the algorithm's own definition of 'interpretable systematic error', not an independent validation against the paper's performance-based definition of systematic error.

  1. self definitional [Section 4.3 'Finding Interpertable Systematic Errors', evaluation paragraph; Eq. (4)]
    "A patch, predicted to be an interpretable systematic error, is verified to be so only if the evaluator finds all of the following conditions to hold: (i) the query patch image depicts a concept different from cj and is coherent to human understanding, (ii) the nearest neighbour patches depict the same human-coherent concept, (iii) the caption for the query patch adequately represents the content of the query patch and its nearest neighbours."

    The human ground truth is constructed from the same three ingredients as the algorithm's scoring function Omega(p) = 1 iff sigma1 + sigma2 - sigma3 >= alpha. Condition (i) is exactly low sigma3 (the patch does not represent cj), condition (ii) is high sigma1/sigma2 (conceptual linkage with neighbours), and condition (iii) is the caption quality that feeds the language-space similarities. Hence Table 2's 'accuracy' measures human-algorithm agreement on the algorithm's own operational definition of an interpretable systematic error, not whether the discovered slices satisfy the paper's Section 2 definition of a systematic error as a slice with significantly lower model performance. The evaluation is self-consistent rather than externally validated.

  2. other [Results paragraph, Section 4.3 (ACDC discussion)]
    "In ACDC, patches do not contain systematic errors (verified by our human evaluation). In fact, the patches only contain random errors that are not human-coherent."

    The claim that ACDC contains no systematic errors is verified by the same human evaluation that uses the algorithm's criteria. The high ACDC accuracies (e.g., 84.21% and 93.54%) are therefore dominated by the algorithm agreeing with itself about true negatives. No external evidence—such as SSM error rates on the discovered slices compared with random slices or class-average baselines—is supplied, so the negative result is also self-referential.

full rationale

The systematicity score Omega (Eq. 4) is not fitted to the human labels: alpha is fixed at 0.35 and the embeddings come from pretrained CLIP/BLIP-2, so the pipeline is not circular in the fitted-input sense. However, the paper's quantitative validation of the central claim—Table 2 'accuracy in assessing interpretable systematic errors'—is self-referential. The human evaluators are asked to verify exactly the three conditions the scoring function already implements: the patch does not show cj (low sigma3), the nearest neighbors are conceptually linked (high sigma1/sigma2), and the BLIP-2 caption is coherent. Thus the reported 95.23% and similar figures measure agreement with the algorithm's own operational definition of 'interpretable systematic error,' not agreement with the Section 2 definition of a systematic error as a slice with significantly lower model performance. No SSM error-rate comparison, random-slice baseline, or oracle validation is supplied. The ACDC 'no systematic errors' claim is likewise verified by the same self-referential human study. A minor self-citation (Gandhi et al. 2020 as an SVM example) is not load-bearing. Overall, the zero-shot pipeline retains independent content, but the central evaluation partially reduces to the algorithm's own definition, giving a score of 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The pipeline rests on four domain assumptions that are stated in the text but not independently validated: the detector defines truth for precision errors, embedding and caption similarity captures human-interpretable coherence, a fixed IoU threshold defines ground truth for detector evaluation, and human majority opinion defines validity for systematicity. There are no free parameters fitted to the reported accuracy numbers, though alpha, q, patch size, and detector thresholds are hand-set.

free parameters (5)
  • alpha (threshold in Eq. 4) = 0.35
    Set empirically by observing cosine similarity ranges in each embedding space (Section 4.3); not chosen on a hold-out set and no sensitivity analysis is reported for the main numbers.
  • q (nearest-neighbor count) = 3
    Paper sets q = 3; ablations show bicycle-class accuracy drops for q = 5 and q = 7, meaning results are sensitive to this choice.
  • minimum patch size a = 60 x 60
    Padches are only used if they meet this size (Section 4.1); ablations at 40 and 80 show accuracy varies by several points.
  • IoU ground-truth threshold = 0.7
    Supplement Alg. 1 and 2 define a patch as genuinely containing the class only if IoU > 0.7; this is the reference for the reported detector accuracies.
  • detector thresholds = GroundingDINO box 0.35, text 0.25; Owl-ViT min sim 0.25
    Used in Section 4.2 to decide whether a patch contains the class; taken from pretrained config but directly sets the error set.
assumptions (5)
  • domain assumption A patch with no detected bounding box for class c_j does not contain c_j.
    Stated in Section 3.2 as the decision rule for precision-error identification; the foundation detector's empty output is treated as ground truth for label-free operation.
  • domain assumption Cosine similarity in CLIP and caption-embedding spaces measures human-interpretable conceptual coherence.
    Sections 3.3.2 and 3.3.3 compute sigma1 and sigma2 from pretrained embeddings and assume high similarity implies shared concept.
  • domain assumption IoU > 0.7 between a patch and the ground-truth class map is the reference for true and false error classification.
    Supplement Alg. 1 and 2 use this threshold to benchmark the detector's precision-error accuracy (Table 1).
  • domain assumption Human evaluator majority defines whether a flagged group is an interpretable systematic error.
    Section 4.3 and the supplement rely on 2-3 evaluators agreeing on three subjective conditions as ground truth.
  • standard math Standard k-NN retrieval and thresholding operations are valid for the claims made.
    k-nearest neighbors in an embedding space and cosine-threshold comparisons are standard operations; no nonstandard mathematics is used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic Discovery and Assessment of Interpretable Systematic Errors in Semantic Segmentation." pith.science (2026). https://pith.science/paper/CXDP6OBW

@misc{pith2026241110845,
  author       = {Pith},
  title        = {Pith review of: Automatic Discovery and Assessment of Interpretable Systematic Errors in Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CXDP6OBW}},
  note         = {Machine review of arXiv:2411.10845}
}
read the original abstract

This paper presents a novel method for discovering systematic errors in segmentation models. For instance, a systematic error in the segmentation model can be a sufficiently large number of misclassifications from the model as a parking meter for a target class of pedestrians. With the rapid deployment of these models in critical applications such as autonomous driving, it is vital to detect and interpret these systematic errors. However, the key challenge is automatically discovering such failures on unlabelled data and forming interpretable semantic sub-groups for intervention. For this, we leverage multimodal foundation models to retrieve errors and use conceptual linkage along with erroneous nature to study the systematic nature of these errors. We demonstrate that such errors are present in SOTA segmentation models (UperNet ConvNeXt and UperNet Swin) trained on the Berkeley Deep Drive and benchmark the approach qualitatively and quantitatively, showing its effectiveness by discovering coherent systematic errors for these models. Our work opens up the avenue to model analysis and intervention that have so far been underexplored in semantic segmentation.

Figures

Figures reproduced from arXiv: 2411.10845 by the authors.

Figure 1
Figure 1. Our framework begins with the inference of an SSM w.r.t. a particular semantic class [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative assessment of systematic errors in the BDD dataset. For the “p [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of systematic error discovery in ACDC. Our framework predicts very few false positives and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Precision, recall, and F1-score metrics for precision error identification for “p [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Precision, recall, and F1-score metrics for precision error identification for “p [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Precision, recall, and F1-score metrics for precision error identification for “p [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Precision, recall, and F1-score metrics for precision error identification for “ [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Precision, recall, and F1-score metrics for precision error identification for “ [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Precision, recall, and F1-score metrics for precision error identification for “ [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 11 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  4. [4]

    R.; and Leyton-Brown, K

    d'Eon, G.; d'Eon, J.; Wright, J. R.; and Leyton-Brown, K. 2022. The spotlight: A general method for discovering systematic errors in deep learning models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency

    work page 2022

  5. [5]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  6. [6]

    ETH VIS Group . 2021. BDD100K Model Zoo . https://github.com/SysCV/bdd100k-models

    work page 2021

  7. [7]

    Eyuboglu, S.; Varma, M.; Saab, K.; Delbrouck, J.-B.; Lee-Messer, C.; Dunnmon, J.; Zou, J.; and R \'e , C. 2022. Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960

    arXiv 2022

  8. [8]

    Feng, D.; Haase-Sch \"u tz, C.; Rosenbaum, L.; Hertlein, H.; Glaeser, C.; Timm, F.; Wiesbeck, W.; and Dietmayer, K. 2020. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3): 1341--1360

    2020

Show all 33 references
  1. [9]

    K.; Shabari Nath, P.; and Chouhan, R

    Gandhi, H. K.; Shabari Nath, P.; and Chouhan, R. 2020. Image glossiness from curvelet features using SVM-based classification. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), 1--6

    2020

  2. [10]

    Guillory, D.; Wang, D.; Ishizaka, S.; Kozuka, K.; and Darrell, T. 2022. ErrorAug: Making Errors to Find Errors in Semantic Segmentation

    2022

  3. [11]

    Jain, S.; Lawrence, H.; Moitra, A.; and Madry, A. 2022. Distilling model failures as directions in latent space. arXiv preprint arXiv:2206.14754

    2022 arXiv

  4. [12]

    Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597

    2023 arXiv

  5. [13]

    Li, J.; Li, D.; Xiong, C.; and Hoi, S. 2022 a . Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 12888--12900. PMLR

    2022

  6. [14]

    H.; Yatskar, M.; Yin, D.; Hsieh, C.-J.; and Chang, K.-W

    Li, L. H.; Yatskar, M.; Yin, D.; Hsieh, C.-J.; and Chang, K.-W. 2019. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557

    2019 arXiv

  7. [15]

    H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-N.; et al

    Li, L. H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-N.; et al. 2022 b . Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10965--10975

    2022

  8. [16]

    Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Li, C.; Yang, J.; Su, H.; Zhu, J.; et al. 2023. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499

    2023 arXiv

  9. [17]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

    2021

  10. [18]

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976--11986

    2022

  11. [19]

    Marufur Rahman, Q.; S \"u nderhauf, N.; Corke, P.; and Dayoub, F. 2021. FSNet: A Failure Detection Framework for Semantic Segmentation. arXiv e-prints, arXiv--2108

    2021

  12. [20]

    Minderer, M.; Gritsenko, A.; and Houlsby, N. 2023. Scaling Open-Vocabulary Object Detection. arXiv preprint arXiv:2306.09683

    2023 arXiv

  13. [21]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning

    2021

  14. [22]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I.; et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8): 9

    2019

  15. [23]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1): 5485--5551

    2020

  16. [24]

    Sakaridis, C.; Dai, D.; and Van Gool, L. 2021. ACDC : The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2021

  17. [25]

    Sohoni, N.; Dunnmon, J.; Angus, G.; Gu, A.; and R \'e , C. 2020. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339--19352

    2020

  18. [26]

    Wang, P.; Chen, P.; Yuan, Y.; Liu, D.; Huang, Z.; Hou, X.; and Cottrell, G. 2018. Understanding convolution for semantic segmentation. In 2018 IEEE winter conference on applications of computer vision (WACV), 1451--1460. Ieee

    2018

  19. [27]

    Wang, R.; Jiang, H.; and Li, Y. 2023. UPerNet with ConvNeXt for Semantic Segmentation. In 2023 IEEE 3rd International Conference on Electronic Technology, Communication and Information (ICETCI), 764--769. IEEE

    2023

  20. [28]

    Xia, Y.; Zhang, Y.; Liu, F.; Shen, W.; and Yuille, A. L. 2020. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part I 16, 145--161. Springer

    2020

  21. [29]

    Xiao, T.; Liu, Y.; Zhou, B.; Jiang, Y.; and Sun, J. 2018. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), 418--434

    2018

  22. [30]

    Xie, C.; Wang, J.; Zhang, Z.; Zhou, Y.; Xie, L.; and Yuille, A. 2017. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, 1369--1378

    2017

  23. [31]

    Yu, C.; Wang, J.; Peng, C.; Gao, C.; Yu, G.; and Sang, N. 2018. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), 325--341

    2018

  24. [32]

    Yu, F.; Chen, H.; Wang, X.; Xian, W.; Chen, Y.; Liu, F.; Madhavan, V.; and Darrell, T. 2020. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2636--2645

    2020

  25. [33]

    Yuan, L.; Chen, D.; Chen, Y.-L.; Codella, N.; Dai, X.; Gao, J.; Hu, H.; Huang, X.; Li, B.; Li, C.; et al. 2021. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432

    2021 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.