- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes RadTok, a radar representation that combines discrete tokenization, range-aware polar resampling, and an inverse-degradation objective. The goal is to suppress radar clutter while reducing the range-dependent resolution variation of polar radar tensors. The method is evaluated on indoor pose estimation and outdoor detection/occupancy prediction using RT-Pose and K-Radar. The reported results show improvements over task-specific baselines and several adapted representation-learning methods.
Critical strengths
-
The paper explicitly targets non-uniform polar resolution and radar clutter/multipath instead of treating radar as a generic image-like input.
-
RadTok can be inserted before existing downstream backbones without changing the task-specific architecture, which makes the design relatively practical.
-
The paper reports results on indoor pose estimation, outdoor 3D detection, and occupancy prediction, with ablations and multiple random seeds, which gives a reasonably broad empirical picture.
Critical weaknesses
-
Cross-sensor generalization is not convincingly addressed. Different FMCW waveforms, antenna layouts, virtual array sizes, and angular resolutions can produce very different radar statistics and artifacts. RadTok does not explicitly model these sensor-dependent factors, and the paper provides no strong zero-shot cross-sensor evaluation. This makes it unclear whether the learned tokenizer captures radar-generic structure or simply adapts to the statistics of the training sensors. The paper itself leaves substantially different sensor configurations for future work.
-
The tokenizer is difficult to assess as a reusable representation without released pretrained weights. Since the main claim is a task-agnostic radar tokenizer, reproducibility should ideally include the pretrained tokenizer/checkpoints, not only the architecture and training recipe. I could not find a clear commitment in the manuscript to release pretrained RadTok weights, which makes the claimed reusable representation hard to independently verify or transfer to other radar datasets.
-
The claimed task-agnostic and transferable representation is not really demonstrated. RadTok is pretrained and then jointly fine-tuned with the task loss separately for each downstream task. There is no convincing cross-task or cross-sensor transfer experiment using one frozen tokenizer. Therefore, the experiments mainly show that the same architecture can be retrained for different tasks, which is much weaker than learning a unified transferable radar representation.
-
The inverse-degradation formulation is severely under-constrained. Both the high-resolution representation and the degradation blur are learned from the same noisy input, while supervision only requires their composition to reconstruct that input. Many different high-resolution representations and blur kernels can satisfy this objective, so there is no guarantee that the recovered feature map corresponds to a physically meaningful clean radar signal.
-
The claim of suppressing sidelobes and multipath is not sufficiently supported. LFQ discretization does not inherently distinguish true targets from sidelobes or multipath, and these artifacts can be stable and highly structured rather than “inconsistent noise.” The paper provides no explicit ghost annotations, PSF measurements, or quantitative sidelobe/multipath suppression evaluation to validate this central claim.
-
The gain from resolution-aware spatialization is confounded with interpolation and increased computation. RadTok upsamples the azimuth dimension by as much as –, causing the downstream network to process a substantially larger feature map. The paper does not provide a clean matched-resolution control using simple interpolation or continuous features, making it difficult to tell whether the gain comes from the proposed radar-native representation or simply denser input and more computation. At , latency increases from 5.8 ms to 36.6 ms and memory from 217 MiB to 980 MiB.
-
Resampling cannot recover missing physical angular information by itself. The method interpolates measurements from a fixed-resolution radar grid and then claims to recover fine-grained structure that the raw input lacks. Without higher-resolution radar ground truth or another physical constraint, this is interpolation rather than demonstrated resolution recovery, and the generated high-resolution structures may be task-driven hallucinations.
-
The main pipeline is not actually token-native. Although the paper is framed as “Radar as Tokens,” the main experiments decode the discrete tokens back into a dense high-resolution feature map before feeding them to existing backbones. Direct use of the tokens is only tested in an appendix experiment on one K-Radar detection setting.
-
The experimental breadth is weaker than the “unified” claim suggests. There are only two radar datasets; two of the three downstream tasks come from the same K-Radar dataset. The paper itself leaves substantially different sensor configurations for future work, so cross-sensor generalization is largely untested.
Questions for authors
- Is the same pretrained RadTok tokenizer transferred across tasks or sensors, or is it retrained/fine-tuned separately for each setting? If the latter, what evidence supports the “task-agnostic” and “transferable” claims?
- How does RadTok handle changes in FMCW waveform, antenna layout, virtual aperture, and angular resolution across radar platforms? Do you have any zero-shot cross-sensor results with a frozen tokenizer?
- Since both the high-resolution representation and the degradation blur are learned from the same noisy input, what prevents non-unique or non-physical solutions that still reconstruct the observation well?
- Can you provide quantitative evidence that the tokenizer actually suppresses sidelobes and multipath, rather than simply improving downstream features?
- Have you compared against a matched-resolution baseline using the same range-dependent upsampling but without discrete tokenization or inverse degradation? This seems necessary to isolate where the gain comes from.
- The main experiments decode tokens back into dense features before the downstream backbone. What advantage does the discrete tokenizer provide over a continuous bottleneck under matched capacity and resolution?