- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies human activity recognition through the fusion of event cameras and WiFi CSI. The authors construct the ExWACT dataset, which includes simulated event data converted from WiMANS videos and paired real-world event and WiFi data collected with an event camera. The method uses modality-specific dual-branch encoders, bidirectional cross-attention, and a temporal consistency loss to fuse features from the two modalities.
Critical strengths
(1) The paper explores joint sensing with event data and WiFi and provides both simulated and real paired data, which supports the study of the complementary properties of these two sensing modalities. (2) The real-world data include different lighting conditions, background human motion, and partial occlusion, which gives the dataset some practical relevance, as described in Section 3.1 and Appendix A.
Critical weaknesses
(1) Although event data reduce dense appearance information, they can still preserve recoverable identity-related features, and some studies discuss the reconstruction of recognizable faces from event streams. Therefore, the event representation itself does not provide a sufficient guarantee of anonymization. The paper does not directly evaluate identity recognizability, so the statement in Table 1 that event data require no post-processing for anonymization lacks direct support. At the same time, the paper does not report a complete quantitative comparison among RGB, RGB-WiFi fusion, and event-WiFi fusion under the same conditions. It therefore remains unclear what benefits and costs result from replacing RGB in terms of recognition accuracy and robustness to low light and occlusion. This makes the motivation for choosing the event modality insufficiently supported. (2) The main operations in Section 3.2 include temporal differencing, convolutional encoding, event voxelization, polarity processing, bidirectional cross-attention, and representation-distance constraints. Much of the novelty appears to lie in the combination of event data and WiFi and the integration of these operations for this task, while the independent algorithmic contribution that specifically addresses the heterogeneous characteristics of the two modalities is not sufficiently clear. (3) Figure 4 does not specify the RGB model, training data, or testing protocol. The RGB predictions also include a rotation class that does not appear in the class list of the real-data branch. It is therefore necessary to confirm whether the two systems use the same task definition. (4) In Figure 14, the diagonal entries of the confusion matrix for THAT on the real-data branch sum to 225 out of 315 samples, which corresponds to an accuracy of approximately 71.43%, while Table 3 reports 78.73%. In the same figure, DECOLLE gives 307/315, or 97.46%, which is consistent with Table 3. If Figure 14 uses a different test set or model version, this difference should be clearly stated. (5) The two projections in Eq. 6 and Eq. 13 in the appendix are both learnable functions, and both can output zero to make the consistency loss equal to zero. Under the current definition, the classification objective does not exclude this degenerate solution. The decrease in accuracy after removing this loss supports its empirical usefulness, but it does not directly demonstrate effective temporal alignment.
Questions for authors
(1) What is the basis for choosing the event modality over RGB? Can the authors provide comparisons of RGB-WiFi fusion and event-WiFi fusion under the same data conditions and with matched models, including accuracy and performance under low-light and occlusion conditions? What direct evidence supports the claim that event data require no anonymization? (2) Compared with standard modality encoding, bidirectional cross-attention, and representation alignment, what new design specifically addresses the heterogeneity between event data and WiFi? Does a standard fusion method with the same encoders and training data pool still perform clearly worse? (3) Do the temporal consistency projections share parameters with the classification path, or are there additional constraints that are not described in the paper? How does the actual implementation prevent both projections from outputting zero and losing their alignment effect?