RF-MatID: Dataset and Benchmark for Radio Frequency Material Identification
Xinyan Chen, Qinchun Li, Ruiqin Ma, Jiaqi Bai, Li Yi, Jianfei Yang
TL;DR
RF-MatID introduces the first large-scale, diverse RF dataset for fine-grained material identification, enabling systematic benchmarking and advancing research in RF-based material recognition for embodied AI applications.
Contribution
The paper presents RF-MatID, a comprehensive open-source RF dataset with diverse categories, frequency range, and controlled perturbations, along with a benchmark for evaluating deep learning models.
Findings
State-of-the-art models achieve baseline performance on RF-MatID.
Benchmark reveals robustness challenges under cross-angle and cross-distance shifts.
Systematic frequency analysis aids real-world deployment insights.
Abstract
Accurate material identification plays a crucial role in embodied AI systems, enabling a wide range of applications. However, current vision-based solutions are limited by the inherent constraints of optical sensors, while radio-frequency (RF) approaches, which can reveal intrinsic material properties, have received growing attention. Despite this progress, RF-based material identification remains hindered by the lack of large-scale public datasets and the limited benchmarking of learning-based approaches. In this work, we present RF-MatID, the first open-source, large-scale, wide-band, and geometry-diverse RF dataset for fine-grained material identification. RF-MatID includes 16 fine-grained categories grouped into 5 superclasses, spanning a broad frequency range from 4 to 43.5 GHz, and comprises 142k samples in both frequency- and time-domain representations. The dataset…
Peer Reviews
Decision·ICLR 2026 Poster
1. First large-scale and open-source RF dataset covering a wide frequency range. The scale and diversity of this dataset can well benifit the community of RF-related machine learning-based study. 2. This paper establishes a well-structured benchmark with multiple frequency protocols, data splits, and nine deep learning models, offering a comprehensive evaluation of model robustness. 3. The authors carefully analyzes both frequency and time domain representations. This leads to the insight that f
1. The term "perturbation-aware" is slightly overclaimed. The authors introduce "perturbation" by varying the distance and angle, which is controllable as part of the measuring/sensing technique itself. As the authors mention, the "real-world interference" is not well considered, like electromagnetic noise, mechanical vibrations, etc. 2. One minor suggestion is that some acronyms should be explained before used, like UWB and MMW. Besides, I am not sure whether "mmWave" and MMW mean the same conc
1. Large dataset and effective data preprocessing: This work collected 71k frequency samples and extend them to time-domain representation, which is a good contribution for real-world applications. In addition, the data preprocessing is carefully designed to augment the data sources. 2. Comprehensive benchmark: The benchmark evaluated different protocols for different real-world usage. The domain adaptation evaluation is conducted on reasonable data split strategies, and the hierachical label sp
1. Presentation: The presentation involves too many mechanical material-specific details, while, correspondingly, lacks AI-oriented intuitions, which is not friendly for a broader community. For example, how the details of Eq. (1) contribute to this design. 2. The benchmarked methods are all general methods that originally proposed for other downstream tasks. Is there any material identification specific methods, or frequency process focused methods worth to be includded? 3. The contributions to
- This work addresses a current gap in RF-based material sensing by providing a new dataset and benchmark focused on material identification. - The dataset is relatively large, covers a wide frequency range, and incorporates variations in acquisition conditions. - The dataset includes of both time- and frequency-domain representations, along with several evaluation protocols targeting both in-distribution performance and robustness to angle and distance shifts. - The authors also benchmark sever
- More discussion on the dataset's limitations, potential biases, and cost or practicality of data collection would help contextualize the scope and applicability of the benchmark. - While the dataset is extensive, the paper would benefit from further clarification on the real-world representativeness of the collection setup and whether the acquisition hardware and environments generalize beyond the authors' configuration. - The evaluation focuses primarily on standard deep learning baselines, a
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsWireless Signal Modulation Classification · RFID technology advancements · Advanced Neural Network Applications