Pith. sign in

REVIEW 1 major objections 6 minor 1 cited by

Future Slot Prediction for Unsupervised Object Discovery in Surgical Video

T0 review · 1 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Dynamic future slot prediction, implemented by a temporal slot transformer and a slot merger, lets unsupervised object-centric models find surgical instruments across video frames and sets state-of-the-art results on three surgical…

desk verdict A credible empirical step for object-centric video models, but the headline margin over Slot-BERT depends on a test-set-chosen merge threshold; worth refereeing with that question on the table. read the letter →

arxiv 2507.01882 v2 pith:3G26TLTV submitted 2025-07-02 cs.CV

Guiqiu Liao , Matjaz Jogan , Marcel Hussing , Edward Zhang , Eric Eaton , Daniel A. Hashimoto This is my paper
classification cs.CV
keywords surgicalvideoobject-centriclearningslotattentionunsupervisedsegmentationdynamicinitializationfuturepredictiontemporaltransformerself-supervision
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

This paper claims that unsupervised object-centric models can parse real surgical videos if the slot representation is initialized by a module that predicts what slots will be present in the next frame. The authors introduce the Dynamic Temporal Slot Transformer (DTST), which reads a buffer of past slot embeddings, is trained with masked auto-encoding, and seeds the next frame's slots by appending empty slots at the last position. A slot-merger module then averages or drops redundant slots so the number of active slots varies from frame to frame. On abdominal, thoracic, and cholecystectomy surgical video benchmarks, the method outperforms fixed-count and adaptive-count slot attention baselines, and transfers zero-shot to the EndoVis domain. The paper also states a limitation: extreme instrument overlap remains hard to disentangle, and long uniform tool shafts can split when occluded beyond the 11-second context window.

What carries the argument

The key machinery is the Dynamic Temporal Slot Transformer (DTST) paired with a slot merger, running on top of a standard slot attention encoder-decoder. Slot attention itself is an attention mechanism that groups image features into a fixed number of latent vectors, or slots, each intended to represent an object. DTST is a transformer that operates on slot embedding vectors rather than word tokens, uses temporal positional embeddings, and is trained with random masking of slots; at inference it appends empty slots to the most recent slot buffer and predicts the future slot with a mask at the last position. The slot merger computes cosine similarity between every pair of slots and averages those above a threshold, producing a binary active-slot mask so the number of slots can differ in every frame. The paper credits this slot-count flexibility for avoiding over-segmentation and under-segmentation, and credits latent-space prediction for the low compute overhead of larger context windows.

What would settle it

Train the model with the DTST's future predictions replaced by the previous frame's slots (or by random slots) while keeping the decoder and merger identical; if segmentation and localization metrics do not drop, future slot prediction is not the cause of the gains. A sharper test is to add an explicit loss on the predicted future slots against the actual slots observed at the next frame and check whether the reconstruction-only objective is leaving performance on the table.

Watch

Extended reading notes

Core claim

The central claim is that dynamic future slot prediction improves unsupervised object discovery in surgical video. Given a sequence of slot embeddings from past frames, DTST predicts the slot initialization for the next frame, and the slot merger adjusts the active slot count before the competitive slot-attention mechanism runs. This makes the model both a next-slot predictor and a temporal reasoner, and the authors report that this combination improves segmentation and localization over Slot-BERT, AdaSlot, and other video object-centric baselines on three surgical domains plus a zero-shot transfer set. The gains are largest on the smallest dataset (thoracic), and the model runs at 5.6 ms per frame with a 20x larger context window still under 100 ms, which the paper ties to latent-space temporal reasoning and a dynamic slot count.

Load-bearing premise

The DTST learns to predict future slots only through the frame reconstruction loss; there is no direct loss on the predicted future slots, so the benefit of dynamic initialization depends on the reconstruction objective providing enough gradient signal for physically plausible future slot predictions.

Editorial extensions

If this is right

  • Surgical video can be parsed into object masks without instrument annotations at inference time, using only self-supervised training.
  • Users no longer need to choose a fixed slot count per dataset; the merger sets the active number of slots per frame.
  • Latent-space future slot prediction keeps per-frame cost around 5.6 ms, making real-time downstream use feasible.
  • Zero-shot transfer to EndoVis indicates that representations learned on one surgical domain can support another.
  • On the abdominal dataset, localization accuracy approaches supervised levels, suggesting unsupervised pre-training can substitute for part of the supervision pipeline.

Reading between the lines

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

  • Beyond the paper, the same dynamic-slot mechanism is a natural fit for other open-set video domains where objects appear and disappear, such as robotic manipulation or driving scenes, since it removes the fixed slot-count assumption.
  • The cosine-similarity threshold acts as a free hyperparameter; a testable extension is to make it adaptive per video or to learn it with a small supervised signal.
  • Because the largest gains appear on the smallest dataset, a strong hypothesis worth testing is that future-slot prediction acts as a data-efficient inductive bias that helps most when labeled or diverse training data are scarce.
  • The paper's noted failure mode at >11 seconds suggests a concrete stress test: run the model on clips longer than the training context to see whether splitting of occluded tools is caused by context length or by the slot-merger threshold.
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

1 major / 6 minor

Summary. The paper proposes a Dynamic Temporal Slot Transformer (DTST) module for unsupervised object discovery in surgical video. The method combines slot attention with a masked temporal transformer that predicts future slot initializations, a slot-merger that adaptively removes or merges redundant slots via a cosine-similarity threshold, and a two-stage training procedure using only the reconstruction loss of Eq. (1). The authors report state-of-the-art results on abdominal, thoracic, and cholecystectomy datasets, plus zero-shot transfer to EndoVis, with ablations supporting each component and a runtime of 5.6 ms per frame. The core empirical claim is that this dynamic future-slot scheme outperforms fixed-slot and adaptive-slot baselines on real surgical videos.

Significance. If the empirical claims hold, this is a meaningful advance for unsupervised object-centric learning in a challenging real-world video domain. The paper addresses an important limitation of fixed-slot approaches, evaluates across multiple surgical datasets with standard metrics, includes zero-shot transfer, ablations, and public code. The computational efficiency claim (5.6 ms/frame) is also practically relevant. The main strength is the breadth of evaluation and the clear ablation evidence that the proposed components, especially the slot merger and future-slot prediction, contribute to the reported gains. However, the central SOTA claim depends on a hyperparameter whose selection procedure is not specified, and the reported margins are not stable across reasonable values of that hyperparameter. The contribution is potentially significant, but the current evidence for the headline comparison is weakened by this issue.

major comments (1)
  1. [Section 2.4 / Fig. 1d] The state-of-the-art claim in Table 1 appears to depend on the cosine-similarity threshold used by the slot merger, but the manuscript does not state how the threshold was selected. Table 3 reports sensitivity on what appears to be the same test set: at threshold 0.80, abdominal mBO-V is 48.9, exactly matching Slot-BERT's 48.9, and at thresholds 0.95 and 0.99 CorLoc drops to 70.3 and 70.2, below Slot-BERT's 70.7. The main results are presumably obtained at 0.90, but this choice is not justified by any validation protocol. Please report a threshold chosen on a held-out validation split (or via cross-validation) and show the comparison to Slot-BERT across thresholds with the selection rule stated. Without this, the reported margins in Table 1 may be inflated by test-set selection.
minor comments (6)
  1. [Section 2.4] The claim that DTST is 'trained both for temporal reasoning and for predicting the optimal future slot initialization' is only supported indirectly: the sole supervision is the reconstruction loss in Eq. (1), and the ablation in Table 2 shows that the X-slot component helps overall, but no direct measure of future-slot prediction accuracy is reported. Adding an analysis of predicted-slot quality (e.g., IoU between predicted and actual slots) would make the mechanism easier to interpret.
  2. [Section 3.2, Table 2] The ablation table uses checkmarks without a key, and the first row does not align visually with the column headers, making it difficult to map each configuration to the reported metrics. Please reformat for readability.
  3. [Section 3.2, Fig. 2] The axis labels and legend in Fig. 2 are very small and partially overlap the data; the caption also does not explain what the bracketed numbers in panels (a) and (b) represent. Please enlarge fonts and clarify.
  4. [Section 3.1] The 'zero-shot' evaluation uses EndoVis, which is also an abdominal surgery dataset; this is zero-shot only in the sense of not being used for training, not in the sense of a different surgical domain. Please clarify this in the text to avoid overstating the generality.
  5. [Section 3.2] The statistical significance statement ('one tailed t-test with Bonferroni correction') does not specify the number of comparisons or whether all metrics and datasets were corrected. Please provide this detail.
  6. [Section 2.4] The slot-merger's drop-path mechanism and the exact merging procedure (e.g., whether merging is iterative, how the binary mask is derived from the threshold) are not described in enough detail for replication. Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reconstruction loss, slot attention, mask readout, and cosine merge are independently defined; Slot-BERT is a baseline, not a load-bearing premise; the threshold sensitivity table is a reproducibility concern, not a circular reduction.

full rationale

The paper's derivation chain is: (i) slot attention produces slots from image features under a reconstruction loss (Eq. 1); (ii) a DTST is trained with the same reconstruction objective and a masking scheme to predict future slots; (iii) a slot merger uses cosine similarity (Eq. 4) to adjust slot count; and (iv) output masks are evaluated against external instrument annotations. Each stage is defined independently, and no reported quantity is an input to its own definition. The only self-citation, Slot-BERT [11], appears as a baseline and architectural contrast ('Unlike previous methods including Slot-BERT [11], which use all available slots for reconstruction, we employ a slot merger'); it is not used to justify the correctness of DTST or to foreclose alternatives. The cosine-similarity threshold is a hyperparameter; Table 3 reports sensitivity on the evaluation data, and the paper's statement that 'thresholds above 0.80 yield stable performance' is an empirical observation rather than a definitional reduction. The reported abdominal numbers in Table 1 correspond to the 0.90 row of Table 3, but the paper does not state how the final threshold was selected, so whether it was tuned on the test set is not established from the text. Moreover, several thresholds above 0.80 beat Slot-BERT, so the SOTA margin is not solely an artifact of selecting the single maximum row. The indirect supervision of DTST future slots through L_recon is a training-objective gap and a correctness concern, not a circularity. Overall, the central comparison is to prior work on external benchmarks, and the derivation is self-contained relative to its stated losses and equations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on standard slot attention components and domain-specific assumptions about surgical video. The main hand-tuned component is the merging threshold, which is swept and selected on the test data. No new physical or conceptual entities are introduced.

free parameters (3)
  • slot similarity threshold = 0.90 (selected via sweep in Table 3)
    Threshold used to decide whether two slots merge. The paper sweeps 0.70-0.99 and adopts 0.90; the main results in Table 1 match the 0.90 row, but no independent validation split is described.
  • initial slot count K = not stated (likely 11)
    The number of slots before merging is a fixed constant; the paper compares Slot-BERT with 5 and 11 slots in Fig. 2b, implying K=11 for the proposed method, but the exact value is not given.
  • temporal context length T = 5, 7, 11 seconds (at 1 FPS)
    The context window length is varied; the paper reports results for 5-second clips in Table 1 and longer clips in Fig. 2a, so T is a chosen hyperparameter.
assumptions (4)
  • domain assumption Surgical scenes can be decomposed into object-centric slots representing instruments and anatomical structures
    The whole method relies on the premise that slot attention can meaningfully bind surgical instruments and tissues; the paper motivates this but does not prove it.
  • domain assumption Feature-level reconstruction from a ViT encoder is a sufficient training signal for object discovery
    The training objective is L_recon in Eq. 1 on ViT features; this is standard in slot attention literature (Seitzer et al.), but the paper does not verify that the learned slots correspond to semantic objects in a causal way.
  • domain assumption Cosine similarity between slot embeddings indicates object identity for merging
    Eq. 4 defines merging based on cosine similarity; this assumes slots of the same object have high similarity, which is not guaranteed.
  • ad hoc to paper Masked transformer training on slots yields temporal reasoning transferable to future slot prediction
    The DTST is trained with random masking, and the paper claims the resulting module can predict future slots when given empty inputs. This transfer from masked reconstruction to future forecasting is an assumption; no explicit future-slot loss is formulated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Future Slot Prediction for Unsupervised Object Discovery in Surgical Video." pith.science (2026). https://pith.science/paper/3G26TLTV

@misc{pith2026250701882,
  author       = {Pith},
  title        = {Pith review of: Future Slot Prediction for Unsupervised Object Discovery in Surgical Video},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3G26TLTV}},
  note         = {Machine review of arXiv:2507.01882}
}
read the original abstract

Object-centric slot attention is an emerging paradigm for unsupervised learning of structured, interpretable object-centric representations (slots). This enables effective reasoning about objects and events at a low computational cost and is thus applicable to critical healthcare applications, such as real-time interpretation of surgical video. The heterogeneous scenes in real-world applications like surgery are, however, difficult to parse into a meaningful set of slots. Current approaches with an adaptive slot count perform well on images, but their performance on surgical videos is low. To address this challenge, we propose a dynamic temporal slot transformer (DTST) module that is trained both for temporal reasoning and for predicting the optimal future slot initialization. The model achieves state-of-the-art performance on multiple surgical databases, demonstrating that unsupervised object-centric methods can be applied to real-world data and become part of the common arsenal in healthcare applications.

Figures

Figures reproduced from arXiv: 2507.01882 by the authors.

Figure 1
Figure 1. (a) An overview of the proposed approach using DTST. (b) Direct initialization of slots at time t + 1 from slot predictions at time t. (c) Dynamic slot initialization using DTST. (d) The overall training pipeline. objective is to minimize Lrecon = ∥Xrecon − X∥ 2 (1) which measures the distance between the reconstructed and original features X [17]. Feature-level reconstruction is more robust to variations in scene a… view at source ↗
Figure 2
Figure 2. (a) Quantitative comparison to five state-of-the-art methods on videos of lengths 5, 7 and 11 seconds. (b) Comparison of our dynamic future slot prediction to Slot-BERT with different static slot counts in different domains [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of unsupervised segmentation using dynamic future slot prediction for abdominal and cholecystectomy surgery videos, compared to Slot-BERT with a fixed number of slots (5 and 11 slots). Ground Truth (GT) on instruments is provided as reference [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning Object-Centric Representations in SAR Images with Multi-Level Feature Fusion

    cs.CV 2025-09 conditional novelty 6.0 of 10

    SlotSAR fuses wavelet scattering features with a SAR foundation model's semantic features to make slot attention separate targets from clutter in SAR images, improving segmentation metrics on ATRNet-STAR.

Reference graph

Works this paper leans on

26 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems, NeurIPS (2023)

    Aydemir, G., Xie, W., Güney, F.: Self-supervised object-centric learning for videos. In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems, NeurIPS (2023)

    work page 2023

  2. [2]

    In: Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems, NeurIPS (2022)

    Elsayed, G.F., Mahendran, A., van Steenkiste, S., Greff, K., Mozer, M.C., Kipf, T.: Savi++: Towards end-to-end object-centric learning from real-world videos. In: Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems, NeurIPS (2022)

    work page 2022

  3. [3]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR (2024) 10 G

    Fan, K., Bai, Z., Xiao, T., He, T., Horn, M., Fu, Y., Locatello, F., Zhang, Z.: Adap- tive slot attention: Object discovery with dynamic slot number. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR (2024) 10 G. Liao et al

    work page 2024

  4. [4]

    In: Proceedings of the 36th International Conference on Ma- chine Learning ICML

    Greff, K., Kaufman, R.L., Kabra, R., Watters, N., Burgess, C., Zoran, D., Matthey, L., Botvinick, M., Lerchner, A.: Multi-object representation learning with iterative variational inference. In: Proceedings of the 36th International Conference on Ma- chine Learning ICML. Proceedings of Machine Learning Research (2019)

    work page 2019

  5. [5]

    In: 5th International Conference on Learning Representations, ICLR (2017)

    Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-VAE: Learning basic visual concepts with a constrained vari- ational framework. In: 5th International Conference on Learning Representations, ICLR (2017)

    work page 2017

  6. [6]

    ArXiv preprint (2020)

    Hong, W.Y., Kao, C.L., Kuo, Y.H., Wang, J.R., Chang, W.L., Shih, C.S.: Cholec- Seg8k: a semantic segmentation dataset for laparoscopic cholecystectomy based on Cholec80. ArXiv preprint (2020)

    work page 2020

  7. [7]

    In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Infor- mation Processing Systems, NeurIPS (2023)

    Jiang, J., Deng, F., Singh, G., Ahn, S.: Object-centric slot diffusion. In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Infor- mation Processing Systems, NeurIPS (2023)

    work page 2023

  8. [8]

    In: 2nd International Conference on Learning Representations, ICLR (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: 2nd International Conference on Learning Representations, ICLR (2014)

    work page 2014

Show all 26 references
  1. [9]

    In: The Tenth International Conference on Learning Representations, ICLR (2022)

    Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jon- schkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: The Tenth International Conference on Learning Representations, ICLR (2022)

    2022

  2. [10]

    In: 8th International Conference on Learning Representations, ICLR (2020)

    Kipf, T.N., van der Pol, E., Welling, M.: Contrastive learning of structured world models. In: 8th International Conference on Learning Representations, ICLR (2020)

    2020

  3. [11]

    ArXiv preprint (2025)

    Liao, G., Jogan, M., Hussing, M., Nakahashi, K., Yasufuku, K., Madani, A., Eaton, E.,Hashimoto,D.A.:Slot-BERT:Self-supervisedobjectdiscoveryinsurgicalvideo. ArXiv preprint (2025)

    2025

  4. [12]

    In: Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems, NeurIPS (2020)

    Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. In: Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Syst...

    2020

  5. [13]

    In: The Twelfth International Con- ference on Learning Representations, ICLR (2024)

    Mansouri, A., Hartford, J.S., Zhang, Y., Bengio, Y.: Object centric architectures enable efficient causal representation learning. In: The Twelfth International Con- ference on Learning Representations, ICLR (2024)

    2024

  6. [14]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (1) (2016)

    Pont-Tuset, J., Arbelaez, P., Barron, J.T., Marques, F., Malik, J.: Multiscale com- binatorial grouping for image segmentation and object proposal generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (1) (2016)

    2016

  7. [15]

    OpenAI blog (8) (2019)

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

    2019

  8. [16]

    In: Ad- vances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems (2017)

    Sabour, S., Frosst, N., Hinton, G.E.: Dynamic routing between capsules. In: Ad- vances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems (2017)

    2017

  9. [17]

    In: The Eleventh International Conference on Learning Representations, ICLR (2023)

    Seitzer, M., Horn, M., Zadaianchuk, A., Zietlow, D., Xiao, T., Simon-Gabriel, C., He, T., Zhang, Z., Schölkopf, B., Brox, T., Locatello, F.: Bridging the gap to real-world object-centric learning. In: The Eleventh International Conference on Learning Representations, ICLR (2023)

    2023

  10. [18]

    In: Forty-first International Confer- ence on Machine Learning, ICML (2024)

    Singh, G., Wang, Y., Yang, J., Ivanovic, B., Ahn, S., Pavone, M., Che, T.: Paral- lelized spatiotemporal slot binding for videos. In: Forty-first International Confer- ence on Machine Learning, ICML (2024)

    2024

  11. [19]

    Singh,G.,Wu,Y.,Ahn,S.:Simpleunsupervisedobject-centriclearningforcomplex and naturalistic videos. In: Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems, NeurIPS (2022) Future Slot Prediction for Unsupervised Object...

    2022

  12. [20]

    IEEE Transactions on Medical Imaging (1) (2016)

    Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., De Mathelin, M., Padoy, N.: EndoNet: A deep architecture for recognition tasks on laparoscopic videos. IEEE Transactions on Medical Imaging (1) (2016)

    2016

  13. [21]

    In: Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Sys- tems, NeurIPS (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Sys- tems, NeurIPS (2017)

    2017

  14. [22]

    ArXiv preprint (2019)

    Watters, N., Matthey, L., Burgess, C.P., Lerchner, A.: Spatial broadcast decoder: A simple architecture for learning disentangled representations in VAEs. ArXiv preprint (2019)

    2019

  15. [23]

    In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Sys- tems, NeurIPS (2023)

    Wu, Z., Hu, J., Lu, W., Gilitschenski, I., Garg, A.: SlotDiffusion: Object-centric generative modeling with diffusion models. In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Sys- tems, NeurIPS (2023)

    2023

  16. [24]

    In: Advances in Neural Informa- tion Processing Systems 36: Annual Conference on Neural Information Processing Systems, NeurIPS (2023)

    Zadaianchuk, A., Seitzer, M., Martius, G.: Object-centric learning for real-world videos by predicting temporal feature similarities. In: Advances in Neural Informa- tion Processing Systems 36: Annual Conference on Neural Information Processing Systems, NeurIPS (2023)

    2023

  17. [25]

    In: IEEE/CVF International Conference on Computer Vision, ICCV (2023)

    Zhao, Z., Wang, J., Horn, M., Ding, Y., He, T., Bai, Z., Zietlow, D., Simon-Gabriel, C.,Shuai,B.,Tu,Z.,Brox,T.,Schiele,B.,Fu,Y.,Locatello,F.,Zhang,Z.,Xiao,T.: Object-centric multiple object tracking. In: IEEE/CVF International Conference on Computer Vision, ICCV (2023)

    2023

  18. [26]

    ArXiv preprint (2023)

    Zia, A., Bhattacharyya, K., Liu, X., Berniker, M., Wang, Z., Nespolo, R., Kondo, S., Kasai, S., Hirasawa, K., Liu, B., et al.: Surgical tool classification and localiza- tion: results and methods from the MICCAI 2022 SurgToolLoc challenge. ArXiv preprint (2023)

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.