Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Dynamic Sub-region Search in Homogeneous Collections Using CLIP

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 A text-plus-box query using CLIP with dynamic region detection and an IoU-based distance doubles recall over static grid partitioning in homogeneous underwater image collections.

desk verdict Useful idea, but the headline recall numbers may be inflated by an unreported drop of images with no overlapping segment. read the letter →

arxiv 2506.09506 v1 pith:XOLRIT25 submitted 2025-06-11 cs.MM

Bastian Jäckl , Vojtěch Kloda , Daniel A. Keim , Jakub Lokoč This is my paper
classification cs.MM
keywords sub-regionsearchobjectdetectionmarineimagesCLIPdynamicpartitioninggeometricconstraintsintersection-over-unionknown-itemretrieval
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

Text-only search with a multimodal model fails when users can only describe a scene with abstract labels like 'coral' or 'fish,' because the resulting candidate sets are too large and too similar. This paper investigates whether letting the user also draw a bounding box around the object they mean fixes the problem. Its proposed pipeline replaces fixed image grids with dynamic regions produced by object segmenters, then ranks by a fused score that combines CLIP semantic similarity with a geometric distance between the query box and each candidate region. The experiments on underwater keyframes show the IoU-based fusion roughly doubling recall over static partitioning (R@100 from 29% to 62% on skippable annotations) and even surpassing a theoretical perfect-crop baseline. The authors explicitly frame all results as upper bounds, since queries were evaluated with perfect ground-truth boxes rather than real user clicks.

What carries the argument

The machinery is a two-stage retrieval pipeline. First, an open-set segmentation or detection model (SAM2, WaterMask, or Grounded-SAM2) partitions each image into semantically coherent regions, and CLIP embeds each region. At query time, a user supplies a text description plus a bounding box; the candidate set is restricted to regions whose box has any overlap ($\mathrm{IoU} > 0$) with the query box, then each candidate is scored by a standardized linear combination of CLIP distance and a rectangle distance, $D_{\text{combined}} = (1-\alpha) D_{\text{CLIP}} + \alpha D_{\text{rect}}$ with $\alpha = 0.5$. The rectangle distance that carries the argument is intersection-over-union ($\mathrm{IoU}$), the fraction of the union of two boxes that is overlapped; it combines position, size, and shape into a single number and acts as both a filter and a ranking signal.

What would settle it

Run the pipeline using bounding boxes collected from actual user clicks on the Marine Video Kit keyframes instead of synthetic Gaussian perturbations, and compare R@100 for the IoU-fused Grounded-SAM2 configuration against the static grid; if the advantage shrinks to near zero, the central robustness claim fails. A more direct check is to measure the distribution of IoU between user clicks and the ground-truth annotations and count how many queries fall below the $\mathrm{IoU} > 0$ filter threshold.

Watch

Extended reading notes

Core claim

The central claim is that position information can rescue semantic search in homogeneous collections when it is enforced as a geometric constraint inside the ranking function, not merely used to select candidate regions. The paper shows that dynamic segmentation with CLIP alone gives almost no gain over a static grid, but fusing CLIP's cosine distance with an intersection-over-union distance between rectangles lifts Grounded-SAM2 to 62% R@100 on skippable annotations and 40% on non-skippable ones, versus 29% and 18% for the static grid. IoU is the strongest of four rectangle distances tested; area, shape, and centroid distances help less. IoU-based ranking also beats the theoretical baseline that crops every frame to the ground-truth annotation, because precise geometric alignment is a stronger evidence of a match than high semantic similarity in a homogeneous domain. The paper concludes that a combination of CLIP with an IoU-based distance is especially effective, doubling recall rates, while noting that all numbers assume perfect query boxes.

Load-bearing premise

The method only works if the user's drawn box overlaps the target object and the region detector produces a segment with positive intersection-over-union against that box, because the candidate filter removes every image that fails this overlap test.

Editorial extensions

If this is right

  • Known-item search systems for homogeneous collections (underwater, medical, archival) can roughly double their recall by letting users draw a box instead of forcing a purely textual query.
  • Dynamic region detection by itself is not the source of the gain; the geometric constraint in the ranking score is what separates the winning configurations from the baselines.
  • A ranking based only on IoU overlap with the query box already outperforms a perfect-crop semantic baseline, so in homogeneous domains geometric agreement can carry more information than subtle semantic differences.
  • The method transfers to another domain only if a region detector accurate enough to produce overlapping segments exists there, which the paper identifies as a precondition.

Reading between the lines

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

  • Editorial inference: real user clicks likely have a non-Gaussian error distribution that depends on object size and image complexity, so collecting actual click data would be the decisive test of whether the reported doubling survives deployment.
  • Editorial inference: the same fused-ranking recipe could be applied to other homogeneous domains where CLIP text queries are weak, such as medical imaging or satellite imagery, whenever a suitable open-set segmenter exists.
  • Editorial inference: the observation that geometry-only ranking beats the semantic oracle suggests a cheap reranking step for existing text-image engines: take a short semantic shortlist and reorder it by geometric consistency with a user-drawn box.
  • Editorial inference: substituting CLIP with a different region-aware visual encoder could reduce the observed sensitivity to box perturbations, a hypothesis the paper does not test.
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 proposes dynamic sub-region retrieval for homogeneous visual collections: instead of a fixed static grid, images are segmented by open-set region detectors (SAM2, WaterMask, Grounded-SAM2), regions are embedded with CLIP, and at query time a user supplies text plus a bounding box. Candidate regions are filtered by overlap with the query box and ranked by a fusion of CLIP semantic distance and one of four geometric distances (area, shape, centroid, IoU). Experiments on Marine Video Kit annotations report recall (R@1/10/100/1000) and mean rank for skippable and non-skippable queries, plus perturbation robustness. The headline claim is that CLIP + IoU-based distance doubles recall relative to static grid partitioning.

Significance. If the central claim is sound, the paper demonstrates a useful recipe for known-item retrieval in domains where text queries are under-specified: dynamically detected regions combined with geometric ranking constraints materially improve recall over static grids. The pipeline is not fitted to the test set: CLIP embeddings and region detectors are pre-trained externally, and the geometric distance uses only the user-provided query box, so there is no circularity in the main comparison. The authors also provide a theoretical upper-bound baseline, which is a good reference point. However, the significance of the headline result is currently limited by missing statistical support and by a candidate-pool asymmetry that may inflate the dynamic methods' recall, so the practical value of the claimed doubling is not yet established.

major comments (4)
  1. [§3.2, Step 7 vs. Table 1] The dynamic pipeline ranks only frames that have at least one detected region with IoU(rect_j, b)>0, so frames with no such region are absent from the ranked output. The static grid, by contrast, has at least one grid cell overlapping any query box, so the static baseline is evaluated over the full database while dynamic methods may be evaluated over a reduced pool. The paper never reports the fraction of database frames with nonempty candidate sets per query, nor the number of frames dropped. If many non-target frames are dropped, a target's rank at fixed cutoffs improves mechanically, so the reported 62% vs. 29% R@100 advantage of Grounded+IoU over Static could be an artifact of pool truncation rather than of geometric ranking. The average frame-coverage analysis in §4.1 (70% vs. 51%) does not address this, because coverage per frame is not the fraction of frames with at least one candidate. Please report the per-query pool sizes, or evaluate Static on the same restricted pool, or add a fallback whole-image candidate for frames with no overlapping segment, and re-examine the headline comparisons.
  2. [§4.1] The text uses 'significantly' repeatedly (e.g., 'IoU significantly outperforms Theoretical', 'dynamic sub-region search models ... significantly outperform static grid-based approaches'), but no variance estimates, no query counts, and no hypothesis tests are reported for these comparisons. The only statistical test in the paper (Wilcoxon signed rank for Semantic vs. Static, p=0.169) concerns a comparison that is not the paper's headline. Please report the number of queries, per-query rank distributions or error bars, and paired significance tests (e.g., Wilcoxon) for IoU fusion vs. Static and IoU fusion vs. Theoretical. Without this, the 'doubling recall' conclusion in Section 5 is not statistically supported.
  3. [§4.1, Table 2] The sentence 'Remarkably, even when rankings are based solely on geometric distances, IoU significantly outperforms Theoretical, as shown in Table 2' is not supported by the cited table. Table 2 shows only rectangle-distance results for SAM, Watermask, and Grounded; the Theoretical row is not present. For Grounded IoU in Table 2, R@1000 is 64 (skippable), which is lower than Theoretical's R@1000 of 67 in Table 1. If the intended comparison is between the fused IoU method in Table 1 and Theoretical, the claim needs to be qualified to specific cutoffs (e.g., R@100) and supported by a significance test.
  4. [§5] The central conclusion that 'A combination of CLIP with an IoU-based distance is especially effective, doubling recall rates' does not account for the candidate-pool asymmetry described above. Because the static baseline is evaluated over all frames while dynamic methods are evaluated only over frames with at least one overlapping region, the ratio reported in the conclusion may overstate the improvement. Please re-derive the doubling claim after equalizing the candidate pools or reporting pool sizes.
minor comments (6)
  1. [§3.2] Typo: 'Preproccesing Stage' should be 'Preprocessing Stage'.
  2. [§3.2] The standardization of all distance functions to N(0,1) is stated but not described: are the mean and standard deviation estimated from all candidate distances per query, or from a fixed training set? Please clarify, because the geometric-mean fusion is sensitive to the standardization procedure.
  3. [§3.2] Figure 2 is referenced for the step-by-step pipeline, but Step 7 (candidate selection via IoU>0) is not explicitly labeled in the figure or in the caption; numbering the steps in the figure would improve readability.
  4. [§4] The paper does not state the total number of keyframes in the Marine Video Kit subset used, nor the number of queries in the skippable and non-skippable partitions. This information is needed to interpret the recall metrics and the magnitude of the reported differences.
  5. [§3.2] Only alpha=0.5 is used for fusion, and the text speculates that other alpha values could improve results further. This is a minor omission, but a small alpha-sweep or at least a sentence explaining why alpha=0.5 was chosen would strengthen the paper.
  6. [Tables 1–5] MNR values are reported with inconsistent formatting (e.g., 9320 vs. 7,715). Use consistent comma formatting or a uniform number format.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning; the comparisons are empirical evaluations with fixed pretrained components and acknowledged oracle query boxes.

full rationale

The paper's derivation chain is empirical rather than formal: it combines fixed CLIP embeddings of automatically detected regions with a geometric distance based on the user-supplied query box. No parameter is fitted to the test annotations, the alpha=0.5 fusion weight is fixed a priori, and the region detectors and CLIP are externally pretrained. The central claim that CLIP plus IoU-based distance doubles recall is supported by in-paper Table 1 calculations, not by reusing cited results as premises. The few self-citations (Jäckl et al. [6] for the 5-grid static setting, the upper-bound baseline, and the annotation collection) only describe the experimental setup; they are not load-bearing evidence for the conclusion. The paper itself labels the perfect-box results as theoretical upper bounds. The candidate-pool truncation noted by the skeptic (images with no detected region overlapping the query box produce no candidates) is a potential internal-validity confound, but it is not circular reasoning: the pipeline does not define the retrieval score in terms of the target label, and the dynamic ranking still requires the semantic CLIP match. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a modest set of assumptions: the user's region query is an accurate bounding box, region detectors produce segments overlapping that box, and the fusion weight is fixed at 0.5. No new entities are introduced. The main free parameter is alpha, chosen by hand.

free parameters (1)
  • Fusion weight alpha = 0.5
    Hand-chosen weight balancing CLIP and geometric distances in the combined score (Section 3.2). The paper evaluates only alpha = 0.5 and notes other values could improve results, so the reported gains are tied to this default.
assumptions (4)
  • domain assumption Users query for specific objects or semantically similar regions (Section 2.2).
    The entire retrieval scenario presumes that region-level query specification matches how users think about homogeneous underwater content.
  • domain assumption Human annotation bounding boxes are optimal query regions (Section 4).
    Annotations are used both as query boxes and as the Theoretical cropping baseline; their quality sets the upper-bound nature of all results.
  • ad hoc to paper Candidate regions are selected by IoU(rect, b) > 0 (Section 3.2, Step 7).
    The paper chooses this overlap criterion for both static and dynamic pipelines; it determines which images can even be ranked.
  • ad hoc to paper All distance functions are standardized to N(0,1) before fusion (Section 3.2).
    The standardization choice affects the relative weight of semantic and geometric distances, especially for geometric mean fusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Sub-region Search in Homogeneous Collections Using CLIP." pith.science (2026). https://pith.science/paper/XOLRIT25

@misc{pith2026250609506,
  author       = {Pith},
  title        = {Pith review of: Dynamic Sub-region Search in Homogeneous Collections Using CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOLRIT25}},
  note         = {Machine review of arXiv:2506.09506}
}
read the original abstract

Querying with text-image-based search engines in highly homogeneous domain-specific image collections is challenging for users, as they often struggle to provide descriptive text queries. For example, in an underwater domain, users can usually characterize entities only with abstract labels, such as corals and fish, which leads to low recall rates. Our work investigates whether recall can be improved by supplementing text queries with position information. Specifically, we explore dynamic image partitioning approaches that divide candidates into semantically meaningful regions of interest. Instead of querying entire images, users can specify regions they recognize. This enables the use of position constraints while preserving the semantic capabilities of multimodal models. We introduce and evaluate strategies for integrating position constraints into semantic search models and compare them against static partitioning approaches. Our evaluation highlights both the potential and the limitations of sub-region-based search methods using dynamic partitioning. Dynamic search models achieve up to double the retrieval performance compared to static partitioning approaches but are highly sensitive to perturbations in the specified query positions.

Figures

Figures reproduced from arXiv: 2506.09506 by the authors.

Figure 1
Figure 1. Examples of image partitioning options for two database images: a) static [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pipeline for dynamic region-based image retrieval. The individual steps are described in the main text. visual scenes without requiring task-specific supervision. This allows us to gen￾erate a set of semantically coherent regions {rj} mi j=1 for each image Ii . Each region rj ⊂ Ii is defined by a bounding box rectj = (leftj ,topj , widthj , heightj ). We use these bounding boxes in Step 3, where we construct CLIP em… view at source ↗
Figure 3
Figure 3. Heatmap showing R@100 for (model, rectangle-distance) pairs for skip [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Heatmap showing R@1000 for (model, rectangle-distance) pairs for skip [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [1]

    https://github.com/IDEA-Research/Grounded-SAM-2, ac- cessed: 2025-05-06

    Grounded-sam-2. https://github.com/IDEA-Research/Grounded-SAM-2, ac- cessed: 2025-05-06

    work page 2025

  2. [2]

    In: He, X., Luo, S., Tao, D., Xu, C., Yang, J., Hasan, M.A

    Blažek, A., Lokoč, J., Matzner, F., Skopal, T.: Enhanced signature-based video browser. In: He, X., Luo, S., Tao, D., Xu, C., Yang, J., Hasan, M.A. (eds.) ACM MMM. pp. 243–248. Springer International Publishing (2015). https://doi.org/10. 1007/978-3-319-14442-9_22

    work page 2015

  3. [3]

    In: IJCNN

    Garcia-D’Urso, N.E., Galan-Cuenca, A., Climent-Pérez, P., Saval-Calvo, M., Azorin-Lopez, J., Fuster-Guillo, A.: Efficient instance segmentation using deep learning for species identification in fish markets. In: IJCNN. pp. 1–8 (2022). https://doi.org/10.1109/IJCNN55064.2022.9892945

    arXiv 2022

  4. [4]

    In: ACM MM

    Hinami, R., Matsui, Y., Satoh, S.: Region-based image retrieval revisited. In: ACM MM. p. 528–536. MM ’17, ACM (2017). https://doi.org/10.1145/3123266.3123312

    work page arXiv 2017

  5. [5]

    In: IROS

    Islam, M.J., Edge, C., Xiao, Y., Luo, P., Mehtaz, M., Morse, C., Enan, S.S., Sattar, J.: Semantic Segmentation of Underwater Imagery: Dataset and Benchmark. In: IROS. IEEE/RSJ (2020). https://doi.org/10.1109/IROS45743.2020.9340821

    arXiv 2020

  6. [6]

    In: ArXiv (2025)

    Jäckl, B., Kloda, V., Keim, D.A., Lokoč, J.: Experimental evaluation of static image sub-region based search models using clip. In: ArXiv (2025)

    work page 2025

  7. [7]

    https://doi.org/10.48550/arXiv.2410.17725

    Khanam, R., Hussain, M.: Yolov11: An overview of the key architectural enhance- ments (2024). https://doi.org/10.48550/arXiv.2410.17725

    arXiv doi:10.48550/arxiv.2410.17725 2024

  8. [8]

    arXiv:2304.02643 (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023). https://doi.org/10.48550/arXiv.2304.02643

    arXiv doi:10.48550/arxiv.2304.02643 2023

Show all 29 references
  1. [9]

    https://doi.org/10.1007/s11042-024-20509-z

    Korfhage, N., Mühling, M., Freisleben, B.: Search anything: segmentation-based similaritysearchviaregionprompts.In:MultimediaToolsandApplications(2024). https://doi.org/10.1007/s11042-024-20509-z

    2024 doi

  2. [10]

    In: ICCV

    Lian, S., Li, H., Cong, R., Li, S., Zhang, W., Kwong, S.: Watermask: Instance segmentation for underwater imagery. In: ICCV. pp. 1305–1315 (October 2023). https://doi.org/10.1109/ICCV51070.2023.00126

    2023

  3. [11]

    arXiv preprint arXiv:2303.05499 (2023)

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open- set object detection. arXiv preprint arXiv:2303.05499 (2023). https://doi.org/10. 48550/arXiv.2303.05499

    2023 doi arXiv

  4. [12]

    In: Gurrin, C., Schoeffmann, K., Jónsson, B.Þ., Dang- Nguyen, D., Lokoč, J., Tran, M., Hürst, W., Rossetto, L., Healy, G

    Lokoč, J., Mejzlík, F., Veselý, P., Soucek, T.: Enhanced somhunter for known- item search in lifelog data. In: Gurrin, C., Schoeffmann, K., Jónsson, B.Þ., Dang- Nguyen, D., Lokoč, J., Tran, M., Hürst, W., Rossetto, L., Healy, G. (eds.) Proceed- ings of the 4th Annual on Lifelo...

    2021

  5. [13]

    (6), 3481–3504 (Aug 2023)

    Lokoč, J., Andreadis, S., Bailer, W., Duane, A., Gurrin, C., Ma, Z., Messina, N., Nguyen, T.N., Peška, L., Rossetto, L., Sauter, L., Schall, K., Schoeffmann, K., Khan,O.S.,Spiess,F.,Vadicamo,L.,Vrochidis,S.:Interactivevideoretrievalinthe ageofeffectivejointembeddingdeepmodels:...

    2023 doi

  6. [14]

    In: CVPR

    Mai, L., Jin, H., Lin, Z., Fang, C., Brandt, J., Liu, F.: Spatial-semantic image search by visual feature synthesis. In: CVPR. pp. 1121–1130 (2017). https://doi. org/10.1109/CVPR.2017.125

    2017 doi

  7. [15]

    G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, 14 Jäckl et al. G., Xu, H., Jegou, H., Mairal,...

    2024 doi arXiv

  8. [16]

    In: 2015 IEEE Winter Conference on Applications of Computer Vision

    Pourian, N., Manjunath, B.: Retrieval of images with objects of specific size, loca- tion, and spatial configuration. In: 2015 IEEE Winter Conference on Applications of Computer Vision. pp. 960–967 (2015). https://doi.org/10.1109/WACV.2015.133

    2015 doi

  9. [17]

    https://doi.org/10.48550/ arXiv.2103.00020

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021). https://doi.org/10.48550/ arXiv.2103.00020

    2021 doi arXiv

  10. [18]

    In: CVPR

    Ranasinghe, K., Shukla, S.N., Poursaeed, O., Ryoo, M.S., Lin, T.Y.: Learning to localize objects improves spatial reasoning in visual-llms. In: CVPR. pp. 12977– 12987 (2024). https://doi.org/10.1109/CVPR52733.2024.01233

    2024

  11. [19]

    https://doi.org/10.48550/arXiv.2408.00714

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos (2024). https://doi.o...

    2024 doi arXiv

  12. [20]

    https://doi

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan,F.,Zeng,Z.,Zhang,H.,Li,F.,Yang,J.,Li,H.,Jiang,Q.,Zhang,L.:Grounded sam: Assembling open-world models for diverse visual tasks (2024). https://doi. org/10.48550/arXiv.2401.14159

    2024 doi arXiv

  13. [21]

    TOMM (1) (Dec 2024)

    Sauter, L., Gasser, R., Schuldt, H., Bernstein, A., Rossetto, L.: Performance eval- uation in multimedia retrieval. TOMM (1) (Dec 2024). https://doi.org/10.1145/ 3678881

    2024

  14. [22]

    In: CVPR

    Shlapentokh-Rothman, M., Blume, A., Xiao, Y., Wu, Y., TV, S., Tao, H., Lee, J.Y., Torres, W., Wang, Y.X., Hoiem, D.: Region-based representations revisited. In: CVPR. pp. 17107–17116 (June 2024). https://doi.org/10.1109/CVPR52733. 2024.01619

    2024

  15. [23]

    In: ACM MM

    Smith, J.R., Chang, S.F.: Visualseek: a fully automated content-based image query system. In: ACM MM. p. 87–98. ACM (1997). https://doi.org/10.1145/244130. 244151

    1997 doi

  16. [24]

    In: ACM MMM

    Stroh, M., Kloda, V., Verner, B., Vopálková, Z., Buchmüller, R., Jäckl, B., Hajko, J., Lokoč, J.: Prak tool v3: Enhancing video item search using localized text and texture queries. In: ACM MMM. pp. 326–333. Springer (2025). https://doi.org/ 10.1007/978-981-96-2074-6_39

    2025 doi

  17. [25]

    In: ACM MMM

    Truong, Q.T., Vu, T.A., Ha, T.S., Lokoč, J., Wong, Y.H., Joneja, A., Yeung, S.K.: Marine video kit: A new marine video dataset for content-based analysis and re- trieval. In: ACM MMM. pp. 539–550. Springer (2023). https://doi.org/10.1007/ 978-3-031-27077-2_42

    2023

  18. [26]

    Vadicamo, L., Arnold, R., Bailer, W., Carrara, F., Gurrin, C., Hezel, N., Li, X., Lokoč, J., Lubos, S., Ma, Z., Messina, N., Nguyen, T.N., Peska, L., Rossetto, L., Sauter, L., Schöffmann, K., Spiess, F., Tran, M.T., Vrochidis, S.: Evaluating per- formance and trends in interac...

    2024 doi

  19. [27]

    In: SIGIR

    Xu, H., Wang, J., Hua, X.S., Li, S.: Image search by concept map. In: SIGIR. p. 275–282. SIGIR ’10, ACM (2010). https://doi.org/10.1145/1835449.1835497

    2010

  20. [28]

    Neurocomputing pp

    Xu, S., Zhang, M., Song, W., Mei, H., He, Q., Liotta, A.: A systematic review and analysis of deep learning-based underwater object detection. Neurocomputing pp. 204–232 (2023). https://doi.org/10.1016/j.neucom.2023.01.056 Dynamic Sub-region Search in Homogeneous Collections U...

    2023 doi

  21. [29]

    https://doi.org/10.48550/arXiv.2306.12156 16 Jäckl et al

    Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything (2023). https://doi.org/10.48550/arXiv.2306.12156 16 Jäckl et al. A Additional experimental results In the following tables and figures, we present additional experimental details, wh...

    2023 doi arXiv

Pith tools

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