Pith. sign in

REVIEW 4 major objections 6 minor 54 references

ADAM: Autonomous Discovery and Annotation Model using LLMs for Context-Aware Annotations

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ADAM labels novel object categories from context and visual similarity alone—no class list, no fine-tuning, no retraining.

desk verdict ADAM is a genuinely new class-list-free labeling pipeline, but the headline numbers are inflated by a CLIP-tangled evaluation metric that needs re-scoring before the claims can be trusted. read the letter →

arxiv 2506.08968 v1 pith:4DYAUS4A submitted 2025-06-10 cs.CV

Amirreza Rouhi , Solmaz Arezoomandan , Knut Peterson , Joseph T. Woods , David K. Han This is my paper
classification cs.CV
keywords open-worldobjectdetectionzero-shotclassificationlargelanguagemodelsCLIPembeddingsembedding-labelrepositorycontext-awarepromptingk-nearest-neighborretrievalself-refinement
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

ADAM is a training-free framework for open-world object labeling. Instead of matching against a fixed list of class names, it asks a large language model to propose candidate labels for an unknown object using the known objects in the scene and CLIP-extracted visual properties such as color and texture. Those candidates are stored alongside CLIP image embeddings in an Embedding-Label Repository, and new objects are labeled by retrieving similar embeddings, voting over the retrieved label lists, and reranking with CLIP text similarity. A self-refinement loop then corrects noisy repository labels through visual cohesion checks and k-nearest-neighbor majority voting. On COCO validation boxes ADAM reaches 61.30% Top-1 accuracy, above CLIP's 45.65% and nearly matching BLIP-VQA's 60.18% while avoiding the class list both baselines require; the paper also reports that this contextual route is load-bearing, since accuracy falls to 6.2% in scenes with zero known objects.

What carries the argument

The Embedding-Label Repository (ELR) — a matrix of $N$ CLIP image embeddings, each paired with an LLM-generated list of candidate labels — is the central object. The Context-Aware Object Prediction (COP) module constructs the prompt that generates those candidates from known-object labels, their bounding boxes, the unknown object's box, and CLIP-selected visual characteristics. Label prediction then runs through three operations: FAISS-based cosine similarity search for the top-$k$ embeddings, frequency-based voting over the retrieved label lists (dropping labels appearing in less than half of the lists), and cross-modal reranking where the CLIP text encoder scores each surviving label against the query embedding. Finally, a self-refinement loop computes intra-class cohesion as average pairwise cosine similarity, flags embeddings whose mean in-class similarity falls below the class cohesion score, and relabels flagged instances by majority vote over their eight nearest neighbors. The information-theoretic motivation is the inequality $\mathrm{H}(Y \mid X_1,\dots,X_n) \le \mathrm{H}(Y \mid X_1,\dots,X_{n-1})$: each added piece of context can only reduce uncertainty about the label.

What would settle it

Take a set of COCO validation images that contain no known objects, run ADAM, and measure Top-1 accuracy; the paper predicts about 6.2%, so a result far above or below that value would respectively refute the claim that context is the load-bearing signal or the claim that visual-only inference is near-chance. A second check is to remove the known-object context from the prompt and compare accuracy at $k=250$, where the paper reports a drop from 42.14% to 8.24%.

Watch

Extended reading notes

Core claim

The central claim is that accurate open-world object naming does not require category supervision, a curated vocabulary, or any gradient updates. ADAM builds an Embedding-Label Repository by feeding a context-rich prompt to LLaMA for every unknown object; the prompt contains the labels and bounding boxes of known objects in the image, a text description of the unknown object's texture/color/material obtained by CLIP caption-matching, and the unknown object's own box. The LLM returns 50 candidate labels, which are stored with the CLIP embedding of the cropped object. At inference, the query embedding retrieves the 250 nearest repository embeddings, the associated label lists are pooled with a 50% frequency threshold, and CLIP text-encoder similarity reranks the survivors to select the final label. The paper reports 61.30% Top-1 accuracy on COCO val boxes (vs 45.65% for CLIP and 60.18% for BLIP-VQA), and shows the repository transfers to PASCAL via Faster R-CNN proposals, where ADAM passes BLIP on 14 of 20 classes while using no label list.

Load-bearing premise

The framework assumes the scene contains enough correctly labeled known objects that, together with CLIP-derived visual characteristics, let the LLM guess the unknown object's label; the paper's own Table 2 shows accuracy collapsing from 61% to 6.2% when zero known objects are present.

Editorial extensions

If this is right

  • Object detectors can be extended to novel categories at inference time by running region proposals through ADAM, without retraining the detector or the vision-language backbones.
  • The embedding-label repository grows as new images are processed, so accuracy should improve with accumulated experience (the paper shows a monotonic benefit from more same-class samples).
  • The same repository transfers across datasets: a repository built from COCO labels PASCAL objects when paired with region proposals, outperforming BLIP on 14 of 20 classes.
  • Self-refinement raises Top-1 accuracy by 3.66 points at $k=250$ entirely without supervision, so noisy LLM labels can be corrected by visual cohesion alone.
  • Because only the LLM's suggestions need to be plausible, the system is not limited to a predefined label vocabulary; it can emit labels outside the training distribution.

Reading between the lines

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

  • The strong dependence on known objects (6.2% accuracy with zero knowns) implies ADAM is better understood as a context-completion system than a visual recognizer; feeding it scenes deliberately sparse in known objects would stress-test that characterization.
  • A testable extension is to replace the CLIP-only visual pathway with region-level features from a self-supervised model to see whether the zero-known-object collapse shrinks; the paper does not explore this.
  • The 0.7 CLIP-similarity threshold used to map predicted labels to ground truth rewards synonyms, so exact-match accuracy on a dataset with a dense label vocabulary (e.g., LVIS) could be substantially lower—an evaluation the paper leaves open.
  • ADAM could serve as an active-learning annotator: run it on images with many known objects, keep high-cohesion labels, and add them back to the repository to bootstrap labeling in progressively less annotated scenes.
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

4 major / 6 minor

Summary. The paper introduces ADAM, a training-free pipeline for open-world object labeling. ADAM builds an Embedding-Label Repository (ELR) by using an LLM (LLaMA 3.2) to propose m=50 candidate labels for masked unknown object regions, conditioned on a prompt containing known objects' labels and boxes plus CLIP-extracted visual characteristics; each candidate label list is stored with the region's CLIP image embedding. At inference, a query region's embedding retrieves k=250 nearest neighbors from the ELR, candidate labels are aggregated by frequency (labels below 50% are removed) and reranked by CLIP text-image similarity, and a self-refinement loop re-labels outliers via cohesion analysis and k_refinement=8 NN majority voting. Experiments on COCO validation boxes report 61.30% Top-1 accuracy, above a CLIP baseline (45.65%) and slightly above BLIP-VQA (60.18%), and a PASCAL Faster R-CNN transfer experiment reports average precision comparable to BLIP. The paper claims that ADAM annotates novel categories using only visual and contextual signals, without fine-tuning or a class list.

Significance. If the evaluation is sound, ADAM is a useful contribution to label-free open-world classification: it combines context-aware LLM prompting with a retrieval-based memory and a training-free refinement loop, and it does not require a fixed class list at inference. The paper has strengths: the system is fully training-free, the repository construction is concrete, the ablation study decomposes the contributions of context, majority voting, cross-modal reranking, and self-refinement, and the PASCAL transfer test at least probes cross-dataset behavior. The significance is conditional, however, because the headline accuracy depends on an evaluation metric that shares an embedding model with the method, on hyperparameters selected on the reported validation set, and on an idealized known-object detector.

major comments (4)
  1. [Sec. 4.1, Eqs. (11)–(13)] Section 4.1 defines a correct prediction as CLIP text-embedding cosine similarity ≥ 0.7 to the ground-truth label, while Eq. (12) uses that same CLIP text encoder to rerank ADAM's candidate labels before Eq. (13) selects L_final. Thus the evaluation yardstick is a component of ADAM itself, and the paper does not calibrate the 0.7 threshold against human judgment, does not show that non-matching COCO class pairs fall below 0.7, and does not state explicitly whether the CLIP and BLIP baselines are scored under the identical semantic-mapping protocol. Because the headline result in Table 1 (61.30% vs. 45.65% for CLIP) rests on this metric, please re-score all methods under a single protocol—for example, exact string match after a fixed synonym mapping, or a threshold validated on held-out pairs—and report both exact-match and threshold-based accuracies.
  2. [Sec. 4.4 and Sec. 4.1] Section 4.1 says k=250 was 'empirically chosen according to section 4.4 (Ablation Study)', and Section 3.5 says k_refinement=8 was 'empirically selected'; Section 4.4 and Table 4 perform that selection on COCO validation performance, and the same split is then reported in Table 1. Selecting hyperparameters on the reported validation set makes the 61.30% figure optimistic. Please use a separate development split or nested validation for choosing k, k_refinement, the 50% frequency threshold in Eq. (10), and the 0.7 evaluation threshold, and report sensitivity of the headline results to these choices.
  3. [Sec. 4.1, Table 2] The COCO evaluation uses ground-truth annotations as the 'known objects' ('Ground truths were chosen to limit the interference from false positives that occur using an object detector'), and Table 2 shows that Top-1 accuracy drops to 6.2% when zero known objects are present. Consequently the reported 61.30% is conditional on a perfect known-object detector and on scenes containing sufficient context; it does not measure behavior under the noisy known-object detections that a real open-world system would encounter. Please report an experiment in which known objects come from a real detector (with false positives and localization errors) or a controlled noise injection, and condition accuracy on the number and correctness of known objects.
  4. [Sec. 3.3, Sec. 4.1, Sec. 4.3] Section 3.3 builds the ELR by iterating over every COCO training class, masking each class in turn; hence the repository contains instances of all 80 COCO classes before COCO validation is evaluated, and Section 4.3's PASCAL experiment reuses this repository whose classes include all 20 PASCAL categories. The experiments therefore do not test annotation of categories that are absent from the repository, despite the abstract's claim about 'novel categories.' Please include a held-out-category experiment (e.g., remove a set of classes from the repository and test on those classes only), or rephrase the claim as label-free classification over categories present in the repository.
minor comments (6)
  1. [Sec. 3.1] The entropy inequality H(Y|X1,...,Xn) ≤ H(Y|X1,...,Xn−1) is presented as 'theoretical motivation,' but it only states that conditioning reduces entropy; it does not establish that the specific contextual variables used in the prompt are relevant to the unknown object's label. Please rephrase this as an intuition rather than a formal derivation.
  2. [Sec. 3.4, Eq. (9)] The notation L_similar = {L_i1,j, L_i2,j, ..., L_ik,j} is ambiguous because each L_i,j is already a list; please define L_similar as a collection of label lists before frequency aggregation.
  3. [Sec. 4.2.2] The text says accuracy rises to 58% for scenes with nine or more known objects, while Table 2 reports 58.2%; please make the rounding consistent.
  4. [Sec. 4.3] Please clarify how the known-object labels and boxes are obtained for ADAM when running on Faster R-CNN proposals on PASCAL; the current description only states that the COCO repository is reused.
  5. [Table 4 and Sec. 4.1] Table 4 contains typos ('V oting' in the row label) and Section 4.1 contains 'Emedding-Label Repository'; please proofread the manuscript.
  6. [Eqs. (8) and (16)] The notation 'arg top-k' is nonstandard; please define the ordering and tie-breaking used when retrieving the top-k nearest neighbors.

Circularity Check

2 steps flagged · score 4.0 of 10

Evaluation metric and validation-set hyperparameter choice are entangled with ADAM's own CLIP scoring, inflating the headline Top-1 accuracy.

  1. other [Section 3.4 (Eqs. 11-13) and Section 4.1 evaluation protocol]
    "The cross-modal similarity between the previously computed visual embedding ⃗vu′ of the unknown object and each text embedding ⃗vtext(l) is calculated as: score(l) = cos(⃗vu′, ⃗vtext(l)). The label with the highest similarity score is then selected as the final prediction: Lfinal = arg max l∈Lsorted score(l). ... To address this, predicted labels were mapped to ground truth labels using CLIP's text encoder and cosine similarity. Labels with a similarity of 0.7 or higher to the ground truth were considered correct."

    ADAM's final label is selected by maximizing CLIP visual-text cosine similarity (Eqs. 12-13), and the evaluation credits any predicted label whose CLIP text embedding is at least 0.7 cosine-similar to the ground-truth label. The same CLIP text encoder is therefore both the model's scoring function and the yardstick for correctness. This allows ADAM to be rewarded for labels that are merely CLIP-close to the true class rather than semantically exact, and the CLIP/BLIP baselines are given the closed COCO label list without the same semantic-mapping credit. The 0.7 threshold is arbitrary and unvalidated, so the headline 61.30% versus 45.65% gap may reflect metric asymmetry rather than a genuine annotation advantage.

  2. fitted input called prediction [Section 4.1 and Section 3.5 (k and k_refinement selection)]
    "For the Label Prediction step, we set k=250 to retrieve the top k visually similar embeddings during the similarity search. This value was empirically chosen according to section 4.4 (Ablation Study). ... Throughout all experiments involving repository self-refinement, we set krefinement = 8 for the nearest neighbor voting step. This value was empirically selected to balance label stability and correction strength."

    The COCO validation set is used both to select the hyperparameters k and k_refinement and to compute the final reported accuracy. The ablation in Table 4 is run on the same validation boxes that Table 1 reports, and the selected configuration, k=250 with self-refinement, yields the headline 61.30% Top-1 accuracy. Thus the advertised result is a selected maximum over a small hyperparameter grid rather than an independent prediction on a held-out set. The fitted retrieval neighborhood is then presented as a general capability, making the headline number partly a test-set-fitting artifact rather than an unbiased estimate of open-world labeling performance.

full rationale

The core ADAM pipeline is not a derivation that reduces to its inputs: it constructs an embedding-label repository from LLM-generated candidate labels and CLIP visual embeddings, retrieves neighbors, applies frequency voting, and refines labels via cohesion analysis. The ground-truth label is never an input to the prediction equations, and the method is self-contained against external benchmarks, so this is not a case of self-definitional circularity or a load-bearing self-citation chain. However, two evaluation-related problems warrant a moderate circularity score of 4. First, the correctness criterion uses CLIP text-embedding similarity with a 0.7 threshold, while ADAM's own cross-modal reranker uses the same CLIP text encoder to select its final label; the yardstick and the model share a component, and the baselines are not given the same semantic-mapping credit. Second, the key hyperparameters k and k_refinement are empirically chosen on the validation set that is also used to report the final accuracy, so the headline 61.30% is partly a selection result. These issues inflate or bias the central claim without making the method's internal derivation circular, hence a score of 4 rather than 6 or higher.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of tuned hyperparameters (k, k_refinement, thresholds, m) and on strong empirical assumptions about the reliability of CLIP embedding geometry and the LLM's ability to infer labels from context. No new physical or conceptual entity is introduced beyond the data structure of the repository.

free parameters (5)
  • k = 250
    Number of nearest neighbors retrieved during label prediction; selected via ablation on the validation set (Section 4.4).
  • k_refinement = 8
    Number of neighbors used in self-refinement majority voting; empirically chosen (Section 3.5).
  • frequency_threshold = 50%
    Labels appearing in fewer than half of the retrieved label lists are discarded (Eq. 10).
  • eval_similarity_threshold = 0.7
    CLIP text similarity threshold used to decide if a predicted label matches the ground truth (Section 4.1).
  • m = 50
    Number of candidate labels generated per object by the LLM (Section 3.3.1).
assumptions (4)
  • domain assumption The CLIP embedding space is semantically smooth: visually similar crops have similar embeddings that correspond to similar labels.
    Used throughout, e.g., Eq. 8 and Eq. 14; the method relies on kNN retrieving same-class objects.
  • domain assumption The LLM can produce the correct or synonymous label among its 50 candidates from a text prompt containing known object labels, bounding boxes, and CLIP-extracted visual characteristics.
    This is the core COP module assumption (Section 3.3.1). If the true label is rarely in the candidate lists, downstream voting cannot recover it.
  • standard math The information-theoretic inequality H(Y|X1..Xn) <= H(Y|X1..Xn-1) is the relevant justification for context-based labeling.
    Section 3.1; it is a true inequality but only a motivation, not a guarantee that LLM-generated context is informative.
  • domain assumption Known objects in the scene are correctly labeled and localized.
    In the COCO eval, ground-truth annotations are used as known objects (Section 4.1); in practice a base detector would be noisy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADAM: Autonomous Discovery and Annotation Model using LLMs for Context-Aware Annotations." pith.science (2026). https://pith.science/paper/4DYAUS4A

@misc{pith2026250608968,
  author       = {Pith},
  title        = {Pith review of: ADAM: Autonomous Discovery and Annotation Model using LLMs for Context-Aware Annotations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DYAUS4A}},
  note         = {Machine review of arXiv:2506.08968}
}
read the original abstract

Object detection models typically rely on predefined categories, limiting their ability to identify novel objects in open-world scenarios. To overcome this constraint, we introduce ADAM: Autonomous Discovery and Annotation Model, a training-free, self-refining framework for open-world object labeling. ADAM leverages large language models (LLMs) to generate candidate labels for unknown objects based on contextual information from known entities within a scene. These labels are paired with visual embeddings from CLIP to construct an Embedding-Label Repository (ELR) that enables inference without category supervision. For a newly encountered unknown object, ADAM retrieves visually similar instances from the ELR and applies frequency-based voting and cross-modal re-ranking to assign a robust label. To further enhance consistency, we introduce a self-refinement loop that re-evaluates repository labels using visual cohesion analysis and k-nearest-neighbor-based majority re-labeling. Experimental results on the COCO and PASCAL datasets demonstrate that ADAM effectively annotates novel categories using only visual and contextual signals, without requiring any fine-tuning or retraining.

Figures

Figures reproduced from arXiv: 2506.08968 by the authors.

Figure 1
Figure 1. Generating the Embedding-Label Repository: This figure illustrates the pipeline for generating the embedding-label repository in ADAM. The Context-Aware Object Prediction (COP) module uses known objects’ labels and locations, along with characteristics of the unknown object extracted using CLIP, to construct a context-rich prompt. The prompt is input to an LLM, which generates candidate labels (Lu,j ) for the unknow… view at source ↗
Figure 2
Figure 2. Schematic of the Label Prediction Process for New Unknown Objects: The unknown object Iu′ is processed through a CLIP image encoder to extract its visual embedding ⃗vu′ . Similar embeddings ⃗vsimilar are retrieved from the embedding-label repository, and their associated candidate labels are ranked using frequency-based ranking. Finally, CLIP’s text encoder performs cross-modal reranking to refine the labels, yieldi… view at source ↗
Figure 3
Figure 3. Top-1 accuracy of ADAM for different object categories as a function of the number of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 30 canonical work pages

  1. [1]

    Deep vit features as dense visual descriptors, 2022

    Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors, 2022. URLhttps://arxiv.org/abs/2112.05814

    arXiv 2022

  2. [2]

    A faster secure content-based image retrieval using clustering for cloud.Expert Systems with Applications, 189:116070, 2022

    J Anju and R Shreelekshmi. A faster secure content-based image retrieval using clustering for cloud.Expert Systems with Applications, 189:116070, 2022

    work page 2022

  3. [3]

    Clustering- based real-time anomaly detection—a breakthrough in big data technologies.Transactions on Emerging Telecommunications Technologies, 33(8):e3647, 2022

    Riyaz Ahamed Ariyaluran Habeeb, Fariza Nasaruddin, Abdullah Gani, Mohamed Ahzam Amanullah, Ibrahim Abaker Targio Hashem, Ejaz Ahmed, and Muhammad Imran. Clustering- based real-time anomaly detection—a breakthrough in big data technologies.Transactions on Emerging Telecommunications Technologies, 33(8):e3647, 2022

    work page 2022

  4. [4]

    Cascade r-cnn: Delving into high quality object detection

    Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6154–6162, 2018

    2018

  5. [5]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229. Springer, 2020

    2020

  6. [6]

    Ovarnet: Towards open-vocabulary object attribute recognition

    Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. InCVPR, 2023

    work page 2023

  7. [7]

    Spann: Highly-efficient billion-scale approximate nearest neighborhood search

    Qi Chen, Bing Zhao, Haidong Wang, Mingqin Li, Chuanjie Liu, Zengzhong Li, Mao Yang, and Jingdong Wang. Spann: Highly-efficient billion-scale approximate nearest neighborhood search. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, 10 editors,Advances in Neural Information Processing Systems, volume 34, pages 5199–5212. Curra...

    work page 2021

  8. [8]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PMLR, 2020

    2020

Show all 54 references
  1. [9]

    Yolo- world: Real-time open-vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo- world: Real-time open-vocabulary object detection. InProc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  2. [10]

    A new subspace clustering strategy for ai-based data analysis in iot system.IEEE Internet of Things Journal, 8 (16):12540–12549, 2021

    Zhihua Cui, Xuechun Jing, Peng Zhao, Wensheng Zhang, and Jinjun Chen. A new subspace clustering strategy for ai-based data analysis in iot system.IEEE Internet of Things Journal, 8 (16):12540–12549, 2021

    2021

  3. [11]

    The faiss library.arXiv preprint arXiv:2401.08281, 2024

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre- Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The faiss library.arXiv preprint arXiv:2401.08281, 2024

    2024 arXiv

  4. [12]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal- network.org/challenges/VOC/voc2012/workshop/index.html, 2012

    2012

  5. [13]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. InCVPR, 2023

    2023

  6. [14]

    Open-vocabulary object detection via vision and language knowledge distillation

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. InInternational Conference on Learning Representations, 2022. URLhttps://openreview.net/forum?id=lL3lnMbR4WU

    2022

  7. [15]

    Ow-detr: Open-world detection transformer

    Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. InCVPR, 2022

    2022

  8. [16]

    Morgan kaufmann, 2022

    Jiawei Han, Jian Pei, and Hanghang Tong.Data mining: concepts and techniques. Morgan kaufmann, 2022

    2022

  9. [17]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. InProceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    2017

  10. [18]

    Open-set image tagging with multi-grained text supervision

    Xinyu Huang, Yi-Jie Huang, Youcai Zhang, Weiwei Tian, Rui Feng, Yuejie Zhang, Yanchun Xie, Yaqian Li, and Lei Zhang. Open-set image tagging with multi-grained text supervision. arXiv e-prints, pages arXiv–2310, 2023

    2023

  11. [19]

    LLMs meet VLMs: Boost open vocabulary object detection with fine-grained descriptors

    Sheng Jin, Xueying Jiang, Jiaxing Huang, Lewei Lu, and Shijian Lu. LLMs meet VLMs: Boost open vocabulary object detection with fine-grained descriptors. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=usrChqw6yK

    2024

  12. [20]

    Billion-scale similarity search with GPUs

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535–547, 2019

    2019

  13. [21]

    K. J. Joseph, Salman H. Khan, Fahad Shahbaz Khan, and Vineeth N. Balasubramanian. Towards open world object detection.CoRR, abs/2103.02603, 2021. URL https://arxiv.org/abs/ 2103.02603

    2021 arXiv

  14. [22]

    Clustering-based anomaly detection in multivariate time series data.Applied Soft Computing, 100:106919, 2021

    Jinbo Li, Hesam Izakian, Witold Pedrycz, and Iqbal Jamal. Clustering-based anomaly detection in multivariate time series data.Applied Soft Computing, 100:106919, 2021

    2021

  15. [23]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InICML, 2022. 11

    2022

  16. [24]

    Grounded language-image pre-training

    Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. InCVPR, 2022

    2022

  17. [25]

    Clusterfomer: clustering as a universal visual learner.Advances in neural information processing systems, 36, 2024

    James Liang, Yiming Cui, Qifan Wang, Tong Geng, Wenguan Wang, and Dongfang Liu. Clusterfomer: clustering as a universal visual learner.Advances in neural information processing systems, 36, 2024

    2024

  18. [26]

    Generative region-language pretraining for open-ended object detection

    Chuang Lin, Yi Jiang, Lizhen Qu, Zehuan Yuan, and Jianfei Cai. Generative region-language pretraining for open-ended object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13958–13968, 2024

    2024

  19. [27]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings,...

    2014

  20. [28]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023

    2023

  21. [29]

    Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang

    Muhammad Maaz, Hanoona Abdul Rasheed, Salman H. Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Multi-modal transformers excel at class- agnostic object detection.CoRR, abs/2111.11430, 2021. URL https://arxiv.org/abs/ 2111.11430

    2021 arXiv

  22. [30]

    Simple open-vocabulary object detection

    Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Shai Av...

    2022

  23. [31]

    Clustergan: Latent space clustering in generative adversarial networks

    Sudipto Mukherjee, Himanshu Asnani, Eugene Lin, and Sreeram Kannan. Clustergan: Latent space clustering in generative adversarial networks. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 4610–4617, 2019

    2019

  24. [32]

    Nagy, Patricia A

    William E. Nagy, Patricia A. Herman, and Richard C. Anderson. Learning words from context. Reading Research Quarterly, 20(2):233–253, 1985. ISSN 00340553. URL http://www. jstor.org/stable/747758

    1985

  25. [33]

    Spice: Semantic pseudo-labeling for image clustering.IEEE Transactions on Image Processing, 31:7264–7278, 2022

    Chuang Niu, Hongming Shan, and Ge Wang. Spice: Semantic pseudo-labeling for image clustering.IEEE Transactions on Image Processing, 31:7264–7278, 2022

    2022

  26. [34]

    Data clustering: application and trends

    Gbeminiyi John Oyewole and George Alex Thopil. Data clustering: application and trends. Artificial Intelligence Review, 56(7):6439–6475, 2023

    2023

  27. [35]

    Learning transferable visual models from natural language supervision.CoRR, abs/2103.00020, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision.CoRR, abs/2103.0002...

    2021 arXiv

  28. [36]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  29. [37]

    Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016

    2016

  30. [38]

    Deepdpm: Deep clustering with an unknown number of clusters

    Meitar Ronen, Shahaf E Finder, and Oren Freifeld. Deepdpm: Deep clustering with an unknown number of clusters. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9861–9870, 2022. 12

    2022

  31. [39]

    Enhancing object detection by leveraging large language models for contextual knowledge

    Amirreza Rouhi, Diego Patiño, and David K Han. Enhancing object detection by leveraging large language models for contextual knowledge. InInternational Conference on Pattern Recognition, pages 299–314. Springer, 2025

    2025

  32. [40]

    A review of clustering techniques and developments.Neurocomputing, 267:664–681, 2017

    Amit Saxena, Mukesh Prasad, Akshansh Gupta, Neha Bharill, Om Prakash Patel, Aruna Tiwari, Meng Joo Er, Weiping Ding, and Chin-Teng Lin. A review of clustering techniques and developments.Neurocomputing, 267:664–681, 2017

    2017

  33. [41]

    Stop using the elbow criterion for k-means and how to choose the number of clusters instead.ACM SIGKDD Explorations Newsletter, 25(1):36–42, 2023

    Erich Schubert. Stop using the elbow criterion for k-means and how to choose the number of clusters instead.ACM SIGKDD Explorations Newsletter, 25(1):36–42, 2023

    2023

  34. [42]

    Sparse r-cnn: End-to-end object detection with learnable proposals

    Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni...

    2021

  35. [43]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  36. [44]

    Scan: Learning to classify images without labels

    Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. InEuropean conference on computer vision, pages 268–285. Springer, 2020

    2020

  37. [45]

    Searching for best practices in retrieval-augmented generation

    Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, Ruicheng Yin, Changze Lv, Xiaoqing Zheng, and Xuanjing Huang. Searching for best practices in retrieval-augmented generation. In Yaser Al-Onaizan, Mohit...

    2024

  38. [46]

    Retccl: Clustering-guided contrastive learning for whole-slide image retrieval.Medical image analysis, 83:102645, 2023

    Xiyue Wang, Yuexi Du, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Retccl: Clustering-guided contrastive learning for whole-slide image retrieval.Medical image analysis, 83:102645, 2023

    2023

  39. [47]

    The effects of context on incidental vocabulary learning.University of Hawaii National Foreign Language Resource Center, 20(2):232–245, 2008

    Stuart Webb. The effects of context on incidental vocabulary learning.University of Hawaii National Foreign Language Resource Center, 20(2):232–245, 2008

    2008

  40. [48]

    How well does clip understand texture? InECCV 2022 CVinW Workshop, 2022

    Chenyun Wu and Subhransu Maji. How well does clip understand texture? InECCV 2022 CVinW Workshop, 2022. URLhttps://arxiv.org/abs/2203.11449

    2022 arXiv

  41. [49]

    Hierarchical clustering supported by reciprocal nearest neighbors.Information Sciences, 527:279–292, 2020

    Wen-Bo Xie, Yan-Li Lee, Cong Wang, Duan-Bing Chen, and Tao Zhou. Hierarchical clustering supported by reciprocal nearest neighbors.Information Sciences, 527:279–292, 2020

    2020

  42. [50]

    Open-vocabulary detr with conditional matching

    Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. InEuropean Conference on Computer Vision, 2022

    2022

  43. [51]

    Open-vocabulary object detection using captions

    Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14393–14402, June 2021

    2021

  44. [52]

    Recognize anything: A strong image tagging model.arXiv preprint arXiv:2306.03514, 2023

    Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model.arXiv preprint arXiv:2306.03514, 2023

    2023 arXiv

  45. [53]

    Towards open-set object detection and discovery

    Jiyang Zheng, Weihao Li, Jie Hong, Lars Petersson, and Nick Barnes. Towards open-set object detection and discovery. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3960–3969, 2022. doi: 10.1109/CVPRW56347.2022. 00441

    2022

  46. [54]

    Detecting twenty-thousand classes using image-level supervision

    Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krähenbühl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. InECCV, 2022. 13

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.