acceptodds
Withdrawn from ICLR 2027

MaskTac: Mask-guided Learned Lossy Compression for Sparse Tactile Glove Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tactile sensing provides direct feedback about physical contact for a wide range of perception and interaction applications. However, high-quality force-based tactile datasets remain limited. The widespread use of tactile data also creates a growing need of efficient compression for storage and transmission. To mitagate these gaps, we firstly introduce TacGlove, a tactile glove dataset collected using a customized wearable tactile glove. It contains 4,660 sequences and 781,969 frames, covering 32 object classes and 8 grasp types. The glove records three-axis contact forces from 1,207 taxels distributed across the hand. Such force-based tactile glove data exhibit pronounced spatiotemporal sparsity and strong correlations within contact regions. Based on these characteristics, we then propose MaskTac, a learned lossy codec for sparse tactile glove data. By combining sparsity-aware quantization with a mask-guided contextual entropy model, MaskTac exploits data spatiotemporal sparsity and reduces spatiotemporal redundancy within contact regions for efficient compression. We further establish a benchmark for tactile compression by adapting traditional and learned visual codecs to TacGlove. The benchmark evaluates both reconstruction fidelity and object and grasp classification accuracy. Experiments show that MaskTac outperforms the benchmarked codecs, achieving BD-rate improvement relative to VTM of 34.04% under PSNR, 32.57% under MS-SSIM, and 58.83% under Contact F1. Ablation studies further validate the effectiveness of the proposed design.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Tactile, Tactile Dataset, Lossy Compression, Data Compression

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DH3kOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a large tactile glove dataset, TacGlove, collected using a custom wearable tactile glove. The glove records three-axis contact forces from 1,207 taxels distributed across the hand. It then develops a lossy perceptual coding scheme for the sparse tactile data, called MaskTac, that combines sparsity with contact region information for efficient compression. The central idea is to encode where contact occurs separately from the force values, and use nearby measurements and previous frames to compress those values efficiently. Furthermore, the paper creates a benchmark for tactile compression by adapting visual codecs to TacGlove. It evaluates reconstruction fidelity and object and grasp classification accuracy.

Critical strengths

Valuable tactile dataset. TacGlove provides dense measurements of both normal and shear forces during human grasping across diverse objects and grasp types, making it a useful resource for tactile perception and compression research.

The compression design matches the data well. MaskTac explicitly exploits sparse contact patterns and spatiotemporal correlations by encoding contact locations separately from force amplitudes. The ablations support the effectiveness of this design, showing substantial bitrate increases when mask-based coding or contextual information is removed.

Critical weaknesses

The baselines do not establish the added value of the learned compression model. Comparisons primarily involve general-purpose visual codecs, without a strong conventional sparse predictive compressor or an appropriately adapted prior tactile-trained codec such as TaCo-L. The ablations demonstrate the importance of explicit sparsity handling, but do not isolate how much learned probability prediction improves over simpler alternatives using the same quantization and contact mask. This leaves the central methodological contribution insufficiently established.

The evaluation does not establish practical suitability for online tactile applications. Quantization uses the minimum and maximum of the complete sequence, requiring future measurements before encoding can proceed with the specified step size. Encoding and decoding latency are also unreported. Moreover, reconstruction metrics and sequence-level classification do not establish preservation of brief, functionally important signals such as contact transitions, peak forces, or slip. These gaps limit the evidence that the reported compression gains translate into useful real-time tactile sensing and control.

Questions for authors

Baseline comparison: Have you compared MaskTac with a conventional sparse predictive compressor using the same quantization and contact mask, or with an appropriately adapted tactile-trained codec such as TaCo-L? If such results are available, what gains remain attributable to learned contextual modeling? This would materially clarify the method’s added value.

Online operation: Does the implementation require the complete sequence to determine its quantization range, as the formulation suggests? If an online version has been evaluated, what are its buffering requirements, encoding/decoding latency, and rate–distortion performance?

Reviewer_x8zbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on the compression of force-based tactile glove data. First, the authors introduce TacGlove, a force-based tactile glove dataset containing 4,660 human grasping sequences with three-axis forces measured by 1,207 taxels, aiming to address the lack of large-scale datasets in this domain. Based on this dataset, the authors further propose MaskTac, a lossy tactile codec specifically designed for sparse force maps. The paper also establishes a benchmark by comparing conventional and learned visual/video codecs, evaluating both reconstruction metrics and frozen downstream object/grasp classification performance.

Critical strengths
  1. It addresses the scarcity of tactile glove datasets that simultaneously provide high-density taxels and continuous 3-axis force measurements.
  2. It provides a useful benchmark to study the important problem of compressing dense tactile glove data.
Critical weaknesses

The key weaknesses are (1) Limited technical novelty and (2) Insufficient experimental validation and practicality.

Specifically:

  1. The contribution of the proposed compression method is quite limited. The sparsity-aware quantization is essentially a simple mid-tread uniform scalar quantization, which naturally maps zero values to zero. Therefore, the term “sparsity-aware” appears to be an overstatement. Furthermore, compared with the context models in [1,2], the additions (mask conditioning, temporal history, and residual refinement from the previous frame) are reasonable but incremental domain adaptations, and the paper does not sufficiently justify their novelty beyond these prior works.
  2. The experimental comparison mainly focuses on image and video compression algorithms, but these methods are not well suited for this type of tactile data. For example, the flattening into a taxel-time image used for image-based codecs may inherently destroy the spatial structure of the tactile signals. More appropriate baselines should include: (1) The quantization method used in this work combined with general-purpose lossless compressors, such as zstd, LZMA, or bzip2; (2) Sparse representations, where only non-zero locations and values are stored (e.g., COO format), followed by entropy coding; (3) The same quantizer combined with non-learning context-adaptive arithmetic coding, such as CABAC or CALIC. A comprehensive comparison in terms of both compression performance and computational costs against these baselines is necessary to demonstrate the actual advantages of the proposed method.
  3. The generalization of the codec is insufficiently validated. The dataset is randomly split at the sequence level, so all object categories, grasp types, and presumably all participants appear in both the training and test sets, which may lead to an overly optimistic evaluation. The authors should evaluate the rate-distortion performance on OOD objects, grasp types, or participants that are excluded from codec training.
  4. The Contact F1 metric conflates two different error types and may favor the proposed method. Because MaskTac reconstructs the zero background exactly, its errors are almost exclusively false negatives (weak contacts quantized to zero), whereas conventional codecs such as VTM introduce small reconstruction noise over the dominant non-contact background, producing false positives. Given the strong class imbalance between contact and non-contact taxels, a single F1 value hides this difference. The authors should report precision and recall separately, analyze the sensitivity to the 0.02 threshold, and additionally report reconstruction error computed on contact taxels only.
  5. At low bitrates, MaskTac may discard weak contacts (|x| < Δ/2\Delta / 2), but the paper does not analyze the impact of this information loss. Since weak contact signals may contain important tactile cues for subtle interaction states, the authors should investigate how the proposed quantization strategy affects the preservation of low-magnitude contact information and its downstream task performance.
  6. A potential limitation is that the proposed model may not be suitable for real-time streaming transmission, since the quantization process requires xmaxx_{max} and xminx_{min}​​, which can only be obtained after observing the complete data sequence. This implies that the method is mainly applicable to compressing locally collected and completed datasets, which is inconsistent with the applications emphasized in the introduction, such as remote contact feedback and interactive haptic experiences.
  7. The paper does not provide encoding and decoding speed measurements or comparisons. The authors should report per-frame encoding and decoding time (with hardware specification) and compare it with the baselines to show whether the codec meets the real-time requirement implied by the target applications.
  8. Some formulations are not sufficiently rigorous. For example, the quantization step Δk=r⋅2−k\Delta_k = r·2^{−k} becomes zero when an entire sequence contains no contact (r = 0), making Eq. (2) undefined. In addition, the mask-based scale prior gkg_k in Eq. (5) is not defined, and the purpose of the round() operation in the mean prediction is not explained.

[1] He, Dailan, et al. "Checkerboard context model for efficient learned image compression." 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021.

[2] He, Dailan, et al. "Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

Questions for authors

See weakness.

Reviewer_5uwpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces MaskTac, a learned lossy compression codec tailored for the pronounced spatiotemporal sparsity of high-density tactile glove data. By combining sparsity-aware quantization with a mask-guided contextual entropy model, it perfectly preserves zero-valued backgrounds while efficiently compressing active contact regions. Furthermore, the authors present the large-scale TacGlove dataset and a comprehensive compression benchmark, demonstrating that MaskTac significantly outperforms traditional visual codecs (like VTM) in both reconstruction fidelity and downstream classification tasks, achieving substantial bitrate savings.

Critical strengths

Tackles the massive storage/transmission bottleneck of high-density tactile data, a highly relevant but underexplored problem in embodied AI.

Valuable Dataset Contribution: Introduces TacGlove, a large-scale, dense 3-axis force dataset (1,207 taxels, ~800k frames) with rich grasp and object annotations.

Tailored & Innovative Method: MaskTac is specifically designed for tactile sparsity, uniquely combining sparsity-preserving quantization with a mask-guided contextual entropy model, rather than just adapting visual codecs.

Critical weaknesses

The abstract is overloaded with specific numbers (e.g., 4,660 sequences, 781,969 frames, 34.04%, 58.83%), which obscures the core scientific insights and makes it tedious to read. The overall writing style feels formulaic and heavily relies on AI-like phrasing, making it difficult to quickly dig out the actual novelty and high-level motivation.

The 1st contribution (TacGlove dataset) and the 3rd contribution (Tactile compression benchmark) are essentially overlapping. The benchmark is merely an evaluation protocol built on top of the dataset. These should be consolidated into a single contribution regarding the "Dataset and Benchmarking Framework" to avoid inflating the contribution list.

The framework diagram looks overly complex and resembles an "AI-generated" schematic. It lacks clear technical intuition (e.g., exact tensor shapes, how the context is effectively fused) and fails to convey the core mechanism any better than the text. A diagram should clarify the architecture, not just look visually busy.

Since this work relies heavily on physical human-grasp data collection using a customized wearable glove, the absence of a supplementary video is a major flaw. Without visual evidence of the actual glove hardware, the grasping process, and the raw data streams, it is difficult to verify the physical realism and the claimed "natural human interactions."

The authors arrange tactile data into a "taxel-time image" (1207T 1207×T) for image codec baselines. This completely ignores the physical topology of the hand. The poor performance of visual codecs might simply be due to this naive spatial arrangement rather than their inherent inability to compress tactile data. A fair comparison requires mapping the data to a 2D hand-shaped topology for all image baselines, not just the video ones.

The proposed "sparsity-aware quantization" and mask extraction (Eq. 3) merely separate exact zeros from non-zeros using a simple threshold (>0). While effective for bitrate saving, this is a trivial thresholding operation rather than a learned sparsity-exploiting mechanism. The technical novelty of the "mask-guided" design is somewhat overstated.

The entropy model uses a very lightweight temporal context. It remains unclear if the performance gains over video codecs (like VTM-Inter or DCVC) come from a superior temporal model, or simply because tactile data is mostly zeros (temporal sparsity). Comparing a lightweight autoregressive model with heavy motion-estimation video codecs is slightly "apples-to-oranges."

Questions for authors

See above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.