Conformal Predictions for Human Action Recognition with Vision-Language Models

Bary Tim; Fuchs Cl\'ement; Macq Beno\^it

REVIEW 3 major objections 5 minor 32 references

Conformal Predictions for Human Action Recognition with Vision-Language Models

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

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

Pith's one-line read For human action recognition, conformal prediction on off-the-shelf vision-language models shrinks candidate class sets, and a calibration-only softmax temperature tune shortens their long tail while keeping the coverage guarantee.

desk verdict Temperature tuning on the calibration set voids the conformal guarantee; the paper's central claim is unsupported, though the empirical tail-reduction observation is worth a second look. read the letter →

arxiv 2502.06631 v2 pith:IL7UORTA submitted 2025-02-10 cs.CV cs.AI

Bary Tim , Fuchs Clément , Macq Benoît This is my paper · ORCID

classification cs.CVcs.AI

keywords conformalpredictionhumanactionrecognitionvision-languagemodelstemperaturetuninglong-taildistributioncoverageguaranteezero-shotclassificationhuman-in-the-loop

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a conformal predictor can be layered on top of an off-the-shelf vision-language model for human action recognition, giving each video clip a small set of plausible action classes instead of a single guess. It further claims that tuning the model's softmax temperature on the calibration set shortens the long right tail of conformal set sizes without requiring extra labeled data and without losing the conformal coverage guarantee at level $1-\alpha$. If this is right, human reviewers in settings such as live video monitoring would usually face fewer options and would rarely be handed an extremely long list, which matters because annotation time grows with the number of choices. The carrier of the argument is the LAC (Least Ambiguous set-valued Classifier) nonconformity score $s_i = 1 - y_{i,k^*}$, whose $1-\alpha$ quantile $\hat{q}$ depends on the temperature, and the proposed rule chooses the temperature that minimizes $\hat{q}$ on the calibration set. Experiments on HMDB51, UCF101, and Kinetics400, across several CLIP architectures, show tail-quantile reductions of up to dozens of classes relative to the default temperature.

What carries the argument

The central object is the softmax temperature $\tau$ of the VLM's zero-shot classifier, which controls how sharply the similarity scores are turned into probabilities. The machinery is the LAC score $s_i = 1 - y_{i,k^*}$ for the ground-truth class of a calibration sample; the paper computes the $1-\alpha$ quantile $\hat{q}(\tau)$ of these scores as a function of $\tau$, observes a roughly convex relationship, and sets $\tau^* = \arg\min_\tau \hat{q}(\tau)$. At test time the conformal set is built with the usual LAC threshold: include every class whose softmax score is at least $1 - \hat{q}(\tau^*)$.

What would settle it

Run the full temperature-tuned conformal pipeline on a labeled video dataset, choosing $\tau^*$ and $\hat{q}$ from a calibration fold and measuring how often the true action class appears in the test set across many folds; if the empirical coverage falls below $1-\alpha$ by more than sampling error, the paper's central coverage claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the temperature parameter $\tau$ of a CLIP-style softmax classifier is an effective control knob for the distribution of conformal set sizes in zero-shot human action recognition. Selecting $\tau^* = \arg\min_\tau \hat{q}(\tau)$, where $\hat{q}$ is the $1-\alpha$ quantile of LAC nonconformity scores on the calibration set, reduces the 0.9, 0.95, and 0.975 quantiles of test set sizes compared with the default $1/\tau = 100$, with the largest gains at small $\alpha$ (up to 77 fewer classes in the tail on Kinetics400). The same choice increases the average set size, so the paper presents temperature as a trade-off between mean set size and tail length. The paper asserts that this tuning preserves the split-conformal coverage guarantee because it uses only the calibration set, and it shows the behavior holds across HMDB51, UCF101, Kinetics400, and CLIP encoders from ViT-B/16 to EVA-18B.

Load-bearing premise

The load-bearing premise, asserted in Section 3.2 without a proof or an empirical coverage check, is that choosing the temperature to minimize the calibration-set quantile does not break the conformal coverage guarantee, even though the same calibration data is used both to pick the temperature and to compute the quantile.

Editorial extensions

If this is right

Tail lengths drop: on Kinetics400 the 0.9-, 0.95-, and 0.975-quantiles of conformal set sizes fall by up to 77 classes relative to the default $1/\tau = 100$, with the largest reduction at the smallest $\alpha$.
No extra data is needed: $\tau^*$ is chosen from the same 10-shot calibration set used to compute $\hat{q}$, so the whole pipeline adds no labeled examples beyond the conformal calibration set.
The effect is not architecture-specific: similar tail reductions appear for ViT-B/16, ViT-B/32, ViT-L/14, ResNet50, ResNet101, EVA-8B, and EVA-18B encoders.
There is a price: temperature tuning increases the average conformal set size, so the operating temperature depends on whether annotation time is dominated by typical clips or by the worst tail.
No VLM fine-tuning is required: the method only changes the temperature, so existing zero-shot human action recognition systems can be made more human-in-the-loop friendly with a one-line change.

Reading between the lines

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

The same temperature-tuning mechanism should apply to any softmax-based classifier whose LAC quantile is a convex-looking function of the temperature, so the tail-shrinking effect is not logically tied to vision-language models; testing it on a standard image classifier would isolate the mechanism.
Because the optimal temperature for tail length differs from the one for average set size, a natural extension is to optimize a weighted objective, for example minimizing the 0.95-quantile subject to a budget on the average set size, rather than minimizing the quantile alone.
If human decision time follows Hick's law or a sigmoidal model, the reported tail-quantile reductions should translate into a concrete reduction in worst-case annotation time; a user study measuring reaction times on conformal sets of different sizes would test this directly.

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.

Desk Editor's Note

read the letter

Short version: the paper has a solid empirical observation buried under an unsupported theoretical claim. The temperature tuning step in Section 3.2 selects τ* by minimizing the calibration quantile and then uses the same calibration set to set the threshold. That breaks the exchangeability assumption behind split-conformal LAC coverage. The authors assert the guarantee is preserved (Sections 1, 3.2, 5) but give no proof and, more tellingly, report no coverage numbers anywhere in Section 4. The stress-test note is correct: even in a stylized two-temperature example where both temperature families have identical score distributions, picking the smaller quantile pushes expected coverage below 1−α. So the central claim is not just unproven; it is likely false as stated.

Credit where it is earned: this is the first application of CP to VLM-based HAR, and the experimental work is clean. Multiple datasets, multiple backbones, code released, and the trade-off between mean and tail set size is honestly documented in Figures 5 and 6. The green dots in Figure 3 showing the tail reduction on test data are real evidence that the heuristic does something beyond overfitting the calibration set. That is genuinely useful information.

The soft spot is not minor, but it is fixable. A separate calibration split for selecting τ*, or nested cross-validation, would restore the guarantee, and then reporting empirical coverage on test data would make the paper sound. The 10-shot calibration sets are also small for quantile estimation; that is a secondary issue but worth noting.

I would not cite this in its current form. But it deserves a serious referee: the error is subtle, the empirical results are suggestive, and a good referee could show the authors how to repair it. As written, reject or major revision.

Referee Report

3 major / 5 minor

Summary. This paper applies least-ambiguous set-valued conformal prediction (LAC) on top of zero-shot CLIP-based human action recognition models, with experiments on HMDB51, UCF101, and Kinetics400. The authors observe that conformal prediction reduces average candidate-class counts but that set-size distributions are long-tailed. They propose tuning the softmax temperature on the calibration set by minimizing the 1-alpha quantile of nonconformity scores, and they claim that this preserves the CP coverage guarantee while reducing the tail of set sizes. Results are reported over 40 random splits for several CLIP backbones.

Significance. If the temperature-tuning proposal were valid, it would be a practically useful, annotation-free way to control conformal set sizes for human-in-the-loop action recognition. The empirical scope is a strength: three datasets, several CLIP architectures, 40 random splits, and test-based validation of the tail reduction (green dots in Figure 3). However, the central theoretical claim is not established. The proposed procedure does not fit the standard split-conformal framework, no finite-sample justification is given, and no empirical coverage check is reported. As it stands, the temperature-tuning scheme can at best be regarded as a heuristic, which substantially weakens the paper's contribution to conformal prediction.

major comments (3)

[Section 3.2, Eq. (2)] Selecting tau* = arg min_tau qhat(tau) from the same calibration scores that are then used to compute the conformal threshold makes the nonconformity score function S(.;tau*) depend on the calibration data. The LAC coverage guarantee described in Section 2.1 requires the score function to be fixed before the calibration set is used, because the proof relies on exchangeability of the n calibration scores and the test score generated by that same function. Here that exchangeability argument fails. No alternative finite-sample proof is provided in Sections 3, 4, or 5. This is load-bearing because the abstract, introduction, and conclusion all state that the temperature tuning preserves CP guarantees.
[Section 4, Figures 3-7] No empirical coverage is reported anywhere in the results. Since the same 40 random splits produce both calibration and test sets, an empirical coverage estimate would be essentially free to compute. The green dots in Figure 3 validate only that the chosen tau* approximates the test-data minimizer of tail size; they do not address whether the advertised 1-alpha coverage actually holds. Without either a valid proof or an empirical coverage check, the central claim that guarantees are preserved is unsupported.
[Section 3.2 and Figures 2-6] The minimization underlying tau* is presented over an implicit grid of 1/tau values, but the paper does not specify the grid, the exact quantile estimator (for example, whether the finite-sample correction (1-alpha)(1+1/n) is applied), or the interpolation rule. This makes the reported tau* and the comparisons in Figures 3-6 difficult to reproduce and prevents verification of the finite-sample coverage statement. The claim in Section 3.2 that the relationship between qhat and tau is convex is also stated without proof; Figure 2 shows only a few curves and does not justify a global unique minimizer.

minor comments (5)

[Figure 6 caption] The caption contains the typo "desiredata" and should read "the desired data" or "the user's needs."
[Section 2.3, Eq. (2)] The text says the softmax becomes the argmax operator when tau = 0; strictly, this is true in the limit tau -> 0, since the softmax is undefined at tau = 0.
[Section 3.1] The phrase "10-shot calibration set" is ambiguous: it should be stated explicitly whether 10 refers to total samples or to samples per class, since the calibration size directly affects the finite-sample guarantee.
[Section 2.2] The claim of being the first to employ CP with VLMs for human action recognition is strong and would benefit from a more careful literature check or a softer formulation.
[Section 4] The statement that tail sizes at tau* were "within 0.775 of those at tau_opt" lacks units and context; clarify whether this is a difference in number of classes or a ratio, and over which settings it is averaged.

Circularity Check

1 steps flagged · score 3.0 of 10

Calibration-set temperature selection makes the calibration quantile minimal by construction and is asserted to preserve coverage without proof, but test-set evaluation gives independent support for the tail reduction.

fitted input called prediction [Section 3.2, 'Temperature tuning for tail size reduction'; see also Section 1 and Figure 2]
"To minimize the tail of the conformal set sizes distribution, we select the temperature τ∗ = arg minτ ∈R q̂(τ ). ... This adjustment can be made using only the calibration set, ensuring no additional data cost for conformal predictor calibration, and preserve the guarantees of the CP framework."

The temperature τ* is defined as the minimizer of the calibration quantile q̂(τ), and the same calibration scores then supply the conformal threshold q̂(τ*). Therefore, the statement that τ* minimizes the calibration-set tail is true by construction, not by empirical finding. More importantly, the LAC coverage guarantee described in Section 2.1 requires the score function to be fixed before the calibration set is used; after selecting τ* from the calibration scores, the exchangeability argument underlying the 1−α quantile no longer applies to S(τ*) at test time. The paper asserts that guarantees are preserved but provides no finite-sample proof and reports no test-set coverage.

full rationale

Most of the paper is not circular: applying conformal prediction to an off-the-shelf VLM is a standard pipeline, and the empirical tail-size comparisons are evaluated on test data across multiple architectures and datasets. The one load-bearing self-referential step is the calibration-based temperature selection in Section 3.2, where τ* is defined as the minimizer of the calibration quantile and the same calibration set is then used to set the conformal threshold. That makes the calibration-set improvement definitional and leaves the central claim of a guarantee-preserving procedure unsupported, because the split-conformal coverage guarantee assumes a fixed score function. The test-set validation partially redeems the empirical tail-reduction claim, so the overall circularity is moderate rather than total.

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

The paper introduces one fitted free parameter (temperature tau) and relies on four axioms. The first two are standard conformal assumptions. The third (convexity) is empirically observed but unproven. The fourth is the critical, unjustified assumption that allows data-dependent temperature selection to be used with the standard coverage theorem.

free parameters (1)

softmax temperature tau = selected as argmin of the 1-alpha quantile q_hat(tau) on the calibration set (values differ per dataset/alpha, around…
The temperature is tuned on the calibration set to minimize the quantile of nonconformity scores, and the resulting value is used to compute prediction sets. This is a fitted parameter, not a constant from prior literature.

assumptions (4)

standard math Calibration and test samples are exchangeable (standard split-conformal assumption)
Required for the LAC coverage guarantee stated in Section 2.1.
standard math The LAC nonconformity score s_i = 1 - y_{i,k*} yields valid coverage for a fixed score function
Given in Section 2.1 and cited to [16,17].
domain assumption The function q_hat(tau) is convex, so the minimizing temperature is unique and reliably estimated
Section 3.2 and Figure 2 show the relationship 'appears convex'; no proof is given, and the optimization over a finite grid is treated as global.
ad hoc to paper Tuning tau on the calibration set and then computing q_hat on the same set preserves the 1-alpha coverage guarantee
Asserted in Section 3.2 ('preserve the guarantees of the CP framework') but not derived. This is the key assumption that is likely false under standard split conformal theory.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Conformal Predictions for Human Action Recognition with Vision-Language Models." pith.science (2026). https://pith.science/paper/IL7UORTA

@misc{pith2026250206631,
  author       = {Pith},
  title        = {Pith review of: Conformal Predictions for Human Action Recognition with Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IL7UORTA}},
  note         = {Machine review of arXiv:2502.06631}
}

read the original abstract

Human-in-the-Loop (HITL) systems are essential in high-stakes, real-world applications where AI must collaborate with human decision-makers. This work investigates how Conformal Prediction (CP) techniques, which provide rigorous coverage guarantees, can enhance the reliability of state-of-the-art human action recognition (HAR) systems built upon Vision-Language Models (VLMs). We demonstrate that CP can significantly reduce the average number of candidate classes without modifying the underlying VLM. However, these reductions often result in distributions with long tails which can hinder their practical utility. To mitigate this, we propose tuning the temperature of the softmax prediction, without using additional calibration data. This work contributes to ongoing efforts for multi-modal human-AI interaction in dynamic real-world environments.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

[1]

INTRODUCTION Modern Computer Vision (CV) systems offer high perfor- mances over a wide variety of tasks, surpassing human exper- tise in some cases. However, many applications still rely on Human-In-The-Loop (HITL) frameworks, either to enhance the performance of the underlying CV approach or due to the critical nature of the application, which requires t...

work page Pith review arXiv 2025
[2]

Conformal Predictions Providing reliable confidence estimates for predictions made by deep learning models is essential in many applications

RELA TED WORK 2.1. Conformal Predictions Providing reliable confidence estimates for predictions made by deep learning models is essential in many applications. CP offers a solution by producing sets of potential output classes for a given input, with theoretical guarantees—under mild distributional assumptions—about the inclusion of the true class. The L...

work page
[3]

Experimental settings We conduct our experiments using three video clip datasets: HMDB51 (51 classes) [23], UCF101 (101 classes) [24] and Kinetics400 (400 classes) [20]

METHODS 3.1. Experimental settings We conduct our experiments using three video clip datasets: HMDB51 (51 classes) [23], UCF101 (101 classes) [24] and Kinetics400 (400 classes) [20]. For HMDB51 and UCF101, we use all available samples and split them uniformly at ran- dom into a 10-shot calibration set and a testing set. For Ki- netics400, we apply the sam...

work page
[4]

Red dots mark 1/τ∗, our estimate for minimizing tail size, while green dots indicate the true optimum 1/τopt when it differs

RESULTS AND DISCUSSION Figure 3 shows how the tail of the conformal set size distribu- tion (quantiles at 0.9, 0.95, and 0.975) varies with the inverse temperature 1/τ across datasets and α values. Red dots mark 1/τ∗, our estimate for minimizing tail size, while green dots indicate the true optimum 1/τopt when it differs. Across all settings, tail sizes a...

work page
[5]

CONCLUSION AND FUTURE WORKS In this paper, we demonstrated how integrating Confor- mal Predictors with off-the-shelf Vision-Language Models (VLMs) can significantly reduce the number of possible classes in Human Action Recognition tasks while maintain- ing coverage guarantees. Our findings highlight the influence of the temperature parameter τ on the dist...

work page
[6]

Behavior recognition via sparse spatio-temporal features,

Piotr Doll ´ar, Vincent Rabaud, Garrison Cottrell, and Serge Belongie, “Behavior recognition via sparse spatio-temporal features,” in 2005 IEEE international workshop on visual surveillance and performance eval- uation of tracking and surveillance. IEEE, 2005, pp. 65– 72

work page 2005
[7]

Fast user-guided video object segmen- tation by interaction-and-propagation networks,

Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim, “Fast user-guided video object segmen- tation by interaction-and-propagation networks,” inPro- 0 100 0.000 0.025 0.050 ViT-L/14 Density Kinetics 400 Baseline T emperature tuning 0 20 40 0.0 0.2 UCF101 20 40 0.00 0.05 0.10 HMDB51 0 100 0.00 0.02 0.04 ViT-B/32 Density 0 20 40 0.00 0.05 0.10 20 40 0.0...

work page 2019
[8]

Human-in-the-loop vehicle reid,

Zepeng Li, Dongxiang Zhang, Yanyan Shen, and Gang Chen, “Human-in-the-loop vehicle reid,” in Proceed- ings of the Thirty-Seventh AAAI Conference on Artifi- cial Intelligence and Thirty-Fifth Conference on Inno- vative Applications of Artificial Intelligence and Thir- teenth Symposium on Educational Advances in Artificial Intelligence, 2023, pp. 6048–6055

work page 2023

Show all 32 references

[9]

Surveillance video querying with a human-in-the-loop,

Michael Stonebraker, Bharat Bhargava, Michael Ca- farella, Zachary Collins, Jenna McClellan, Aaron Sipser, Tao Sun, Alina Nesen, K Solaiman, Ganapa- thy Mani, et al., “Surveillance video querying with a human-in-the-loop,” in Proceedings of the Workshop on Human-In-the-Loop Da...

2020
[10]

De- signing decision support systems using counterfactual prediction sets,

Eleni Straitouri and Manuel Gomez Rodriguez, “De- signing decision support systems using counterfactual prediction sets,” in Forty-first International Conference on Machine Learning, 2024

2024
[11]

Conformal prediction sets improve human de- cision making,

Jesse C Cresswell, Yi Sui, Bhargava Kumar, and No ¨el V ouitsis, “Conformal prediction sets improve human de- cision making,” in Forty-first International Conference on Machine Learning, 2024

2024
[12]

Learning transferable visual models from natural lan- guage supervision,

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al., “Learning transferable visual models from natural lan- guage supervision,” in International conference on ma- chine lea...

2021
[13]

Temporal segment networks: Towards good practices for deep ac- tion recognition,

Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool, “Temporal segment networks: Towards good practices for deep ac- tion recognition,” in European conference on computer vision. Springer, 2016, pp. 20–36

2016
[14]

Slowfast networks for video recog- nition,

Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He, “Slowfast networks for video recog- nition,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6202–6211

2019
[15]

Stm: Spatiotemporal and motion en- coding for action recognition,

Boyuan Jiang, MengMeng Wang, Weihao Gan, Wei Wu, and Junjie Yan, “Stm: Spatiotemporal and motion en- coding for action recognition,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 2000–2009

2019
[16]

Vivit: A video vision transformer,

Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid, “Vivit: A video vision transformer,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6836–6846

2021
[17]

Ac- tionclip: A new paradigm for video action recognition,

Mengmeng Wang, Jiazheng Xing, and Yong Liu, “Ac- tionclip: A new paradigm for video action recognition,” arXiv preprint arXiv:2109.08472, 2021

2021 arXiv
[18]

An electronic nose-based assistive diagnostic prototype for lung cancer detection with conformal prediction,

Xianghao Zhan, Zhan Wang, Meng Yang, Zhiyuan Luo, You Wang, and Guang Li, “An electronic nose-based assistive diagnostic prototype for lung cancer detection with conformal prediction,” Measurement, vol. 158, pp. 107588, 2020

2020
[19]

Are foundation models for computer vision good conformal predictors?,

Leo Fillioux, Julio Silva-Rodr ´ıguez, Ismail Ben Ayed, Paul-Henry Courn `ede, Maria Vakalopoulou, Stergios Christodoulidis, and Jose Dolz, “Are foundation models for computer vision good conformal predictors?,” arXiv preprint arXiv:2412.06082, 2024

2024
[20]

On the rate of gain of information,

William E Hick, “On the rate of gain of information,” Quarterly Journal of experimental psychology , vol. 4, no. 1, pp. 11–26, 1952

1952
[21]

Selection from alphabetic and numeric menu trees using a touch screen: breadth, depth, and width,

Thomas K Landauer and Daniel W Nachbar, “Selection from alphabetic and numeric menu trees using a touch screen: breadth, depth, and width,” ACM SIGCHI Bul- letin, vol. 16, no. 4, pp. 73–78, 1985

1985
[22]

29, Springer, 2005

Vladimir V ovk, Alexander Gammerman, and Glenn Shafer, Algorithmic learning in a random world, vol. 29, Springer, 2005

2005
[23]

Least ambiguous set-valued classifiers with bounded error levels,

Mauricio Sadinle, Jing Lei, and Larry Wasserman, “Least ambiguous set-valued classifiers with bounded error levels,” Journal of the American Statistical As- sociation, vol. 114, no. 525, pp. 223–234, 2019

2019
[24]

Ucf101: A dataset of 101 human ac- tions classes from videos in the wild,

K Soomro, “Ucf101: A dataset of 101 human ac- tions classes from videos in the wild,” arXiv preprint arXiv:1212.0402, 2012

2012 arXiv
[25]

to capture the dependence of decision time on task un- certainty, for which the number of classes in the conformal sets could serve as a proxy. This highlights the importance of controlling the tail of the set size distribution, especially for applications such as live video m...
[26]

Dense trajectories and motion bound- ary descriptors for action recognition,

Heng Wang, Alexander Kl ¨aser, Cordelia Schmid, and Cheng-Lin Liu, “Dense trajectories and motion bound- ary descriptors for action recognition,” International journal of computer vision, vol. 103, pp. 60–79, 2013

2013
[27]

Quo vadis, ac- tion recognition? a new model and the kinetics dataset,

Joao Carreira and Andrew Zisserman, “Quo vadis, ac- tion recognition? a new model and the kinetics dataset,” in proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition, 2017, pp. 6299–6308

2017
[28]

Expanding language-image pre- trained models for general video recognition,

Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling, “Expanding language-image pre- trained models for general video recognition,” in Euro- pean Conference on Computer Vision . Springer, 2022, pp. 1–18

2022
[29]

Fine-tuned clip models are efficient video learners,

Hanoona Rasheed, Muhammad Uzair Khattak, Muham- mad Maaz, Salman Khan, and Fahad Shahbaz Khan, “Fine-tuned clip models are efficient video learners,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6545–6554

2023
[30]

Hmdb: A large video database for human motion recognition,

H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “Hmdb: A large video database for human motion recognition,” in 2011 International Conference on Computer Vision, 2011, pp. 2556–2563

2011
[31]

On sequence learning models: Open-loop control not strictly guided by hick’s law,

Rodrigo Pav ˜ao, Joice P Savietto, Jo˜ao R Sato, Gilberto F Xavier, and Andr ´e F Helene, “On sequence learning models: Open-loop control not strictly guided by hick’s law,” Scientific reports, vol. 6, no. 1, pp. 23018, 2016

2016
[32]

Eva- clip-18b: Scaling clip to 18 billion parameters,

Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang, “Eva- clip-18b: Scaling clip to 18 billion parameters,” arXiv preprint arXiv:2402.04252, 2024

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

INTRODUCTION Modern Computer Vision (CV) systems offer high perfor- mances over a wide variety of tasks, surpassing human exper- tise in some cases. However, many applications still rely on Human-In-The-Loop (HITL) frameworks, either to enhance the performance of the underlying CV approach or due to the critical nature of the application, which requires t...

work page Pith review arXiv 2025

[2] [2]

Conformal Predictions Providing reliable confidence estimates for predictions made by deep learning models is essential in many applications

RELA TED WORK 2.1. Conformal Predictions Providing reliable confidence estimates for predictions made by deep learning models is essential in many applications. CP offers a solution by producing sets of potential output classes for a given input, with theoretical guarantees—under mild distributional assumptions—about the inclusion of the true class. The L...

work page

[3] [3]

Experimental settings We conduct our experiments using three video clip datasets: HMDB51 (51 classes) [23], UCF101 (101 classes) [24] and Kinetics400 (400 classes) [20]

METHODS 3.1. Experimental settings We conduct our experiments using three video clip datasets: HMDB51 (51 classes) [23], UCF101 (101 classes) [24] and Kinetics400 (400 classes) [20]. For HMDB51 and UCF101, we use all available samples and split them uniformly at ran- dom into a 10-shot calibration set and a testing set. For Ki- netics400, we apply the sam...

work page

[4] [4]

Red dots mark 1/τ∗, our estimate for minimizing tail size, while green dots indicate the true optimum 1/τopt when it differs

RESULTS AND DISCUSSION Figure 3 shows how the tail of the conformal set size distribu- tion (quantiles at 0.9, 0.95, and 0.975) varies with the inverse temperature 1/τ across datasets and α values. Red dots mark 1/τ∗, our estimate for minimizing tail size, while green dots indicate the true optimum 1/τopt when it differs. Across all settings, tail sizes a...

work page

[5] [5]

CONCLUSION AND FUTURE WORKS In this paper, we demonstrated how integrating Confor- mal Predictors with off-the-shelf Vision-Language Models (VLMs) can significantly reduce the number of possible classes in Human Action Recognition tasks while maintain- ing coverage guarantees. Our findings highlight the influence of the temperature parameter τ on the dist...

work page

[6] [6]

Behavior recognition via sparse spatio-temporal features,

Piotr Doll ´ar, Vincent Rabaud, Garrison Cottrell, and Serge Belongie, “Behavior recognition via sparse spatio-temporal features,” in 2005 IEEE international workshop on visual surveillance and performance eval- uation of tracking and surveillance. IEEE, 2005, pp. 65– 72

work page 2005

[7] [7]

Fast user-guided video object segmen- tation by interaction-and-propagation networks,

Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim, “Fast user-guided video object segmen- tation by interaction-and-propagation networks,” inPro- 0 100 0.000 0.025 0.050 ViT-L/14 Density Kinetics 400 Baseline T emperature tuning 0 20 40 0.0 0.2 UCF101 20 40 0.00 0.05 0.10 HMDB51 0 100 0.00 0.02 0.04 ViT-B/32 Density 0 20 40 0.00 0.05 0.10 20 40 0.0...

work page 2019

[8] [8]

Human-in-the-loop vehicle reid,

Zepeng Li, Dongxiang Zhang, Yanyan Shen, and Gang Chen, “Human-in-the-loop vehicle reid,” in Proceed- ings of the Thirty-Seventh AAAI Conference on Artifi- cial Intelligence and Thirty-Fifth Conference on Inno- vative Applications of Artificial Intelligence and Thir- teenth Symposium on Educational Advances in Artificial Intelligence, 2023, pp. 6048–6055

work page 2023