Spiking Transformer-CNN for Event-based Object Detection

1. The integration of global information from Spiking Transformer with local information from Spiking CNNs was first proposed in [1], rather than in this work.
2. The used or proposed modules in this paper, including SSA, SCB, and both intra-stage and inter-stage spike feature fusion modules, fail to preserve the spiking characteristics of SNNs due to their incorporation of non-spiking computations. Consequently, Spike-TransCNN would be more accurately categorized as a hybrid ANN-SNN model rather than a pure SNN.
3. The paper omits comparisons with other pure SNN models (such as SpikeYOLO [2]) and hybrid models (like EAS-SNN [3] and SpikingViT [4]). Furthermore, the model's mAP performance on GEN1 and 1Mpx datasets is substantially inferior to these state-of-the-art approaches.
4. The motivation for proposing STS is unclear - why is it used in shallow layers instead of SSA?
5. The motivation for proposing SCB is also inadequately explained - on lines 260 and 263, it merely states that it "leverages the local feature extraction capabilities" and "captures local multiscale information". However, this raises questions: Couldn't standard 3x3 and 5x5 convolutions achieve the same objective?
6. The reported energy consumption raises significant concerns. Given that Spike-TransCNN has double the parameters of SFOD with similar firing rates, and extensively employs SPIKE-ELEMENT-WISE ADDITION operations (introducing non-spiking computations) across its modules, the claimed lower energy consumption compared to SFOD requires further justification.
7. The paper contains formatting issues, specifically on line 382 where the Firing Rate and Energy (mJ) are overlapping.

[1] Yao M, Hu J K, Hu T, et al. Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips[C]//The Twelfth International Conference on Learning Representations. [2] Luo X, Yao M, Chou Y, et al. Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection[J]. arXiv preprint arXiv:2407.20708, 2024. [3] Wang Z, Wang Z, Li H, et al. EAS-SNN: End-to-End Adaptive Sampling and Representation for Event-based Detection with Recurrent Spiking Neural Networks[J]. arXiv preprint arXiv:2403.12574, 2024. [4] Yu L, Chen H, Wang Z, et al. Spikingvit: a multi-scale spiking vision transformer model for event-based object detection[J]. IEEE Transactions on Cognitive and Developmental Systems, 2024.

1. Although the hybrid architecture is described in detail, the explanations of specific processes, such as spike-driven token selection and intra- and inter-stage feature fusion, could be clearer. Including pseudo-code or flow diagrams might enhance the reader’s understanding of the model’s operation.
2. While the results on the Gen1 dataset are compelling, it would strengthen the paper to evaluate the model on more diverse datasets, particularly larger or more complex event-based datasets, to demonstrate generalizability.
3. The paper could benefit from a discussion on how Spike-TransCNN compares with hybrid SNN-ANN models, given their potential to balance energy efficiency and performance. This would contextualize the performance and energy efficiency gains of Spike-TransCNN more effectively.
4. While there is an ablation study on some components, further exploration on the impact of various hyperparameters (e.g., number of time steps, membrane potential thresholds) could provide insights into optimizing the architecture for different applications.
5. Consistent terminology, particularly around spiking mechanisms and attention mechanisms, would improve readability. Some abbreviations and terms could be clarified for non-specialist readers.
6. The paper includes visualization results, but providing side-by-side comparisons with other models on challenging scenarios could offer a clearer view of the model’s strengths in handling occlusions and motion.

1. The performance on the 1Mpx dataset (0.250) is significantly inferior to existing methods (0.483), without adequate explanation or analysis.
2. Lack of comprehensive comparisons with recent state-of-the-art SNN detection methods, such as "Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection."
3. The whole architecture primarily transplants the established Transformer-CNN paradigm into the SNN domain, with limited Innovation.
4. The paper exhibits an overreliance on descriptive language while lacking theoretical analysis for the performance improvements of the proposed architecture.

The introduction of this article does not highlight the innovative points. The combination of Transformer and CNN has been proven effective and maturely applied in ANN. The use of LIF neurons for Spiking is not an important innovation point. Furthermore, in the theoretical section, it is evident that both the Transformer architecture and CNN module were designed with reference to previous tasks, which does not constitute an important innovation point to support the paper. The subsequent feature fusion module described two types separately, but the differences between types and the issues addressed were not elaborated in detail. Reading through the methods section, the innovative points and logical description of the methods are similar to the module stacking without highlighting the author's own analysis.