Major Concern:

While this work serves as an early exploration of binary quantization for the occupancy prediction task, several significant limitations are apparent. The study does present two main findings in the context of binary occupancy networks:

• 1×1 binarized convolution introduces only minimal binarization errors as network depth increases.
• Binarized convolution is notably less effective than full-precision convolution at capturing cross-channel feature importance.

However, the proposed strategies to address these issues—namely, enhancing network depth through additional 1x1 binarized layers and adding a per-channel refinement branch—have already been extensively explored in the quantization literature. This repetition limits the novelty of the work, and the paper does not provide fresh insights into the specific task of occupancy prediction. The binary quantization methods introduced here, while perhaps incrementally improving upon existing binary quantization techniques, do not directly tackle the occupancy task itself. Instead, they represent a basic application of binary quantization to occupancy, yielding largely expected results. As such, I believe this work falls short of the standards expected at a top-tier machine learning conference, and the main track may not be the optimal venue for this submission.

Minor Comments:

• This study restricts its exploration to CNN-based approaches for occupancy tasks, without incorporating transformer-based occupancy networks. Transformer models have gained traction and demonstrated superior performance in related fields, and their absence here limits the study’s relevance to current research trends.
• The paper outlines a manually crafted binary network architecture aimed at mitigating quantization-induced errors. However, quantizing both weights and activations from 32-bit to 1-bit introduces significant quantization errors. Without innovations specifically targeted at reducing such errors in this context, the proposed method struggles to demonstrate substantial improvements.
• The speed advantage observed (6.22 FPS versus 7.53 FPS) is marginal, casting doubt on the practical benefits and motivation of the approach. If the goal is to address computational cost and memory constraints in existing occupancy networks, a rigorous analysis of actual speedup and memory reduction should be provided. Alternatively, if the study’s focus is to explore novel quantization techniques for this task, one would expect groundbreaking insights tailored to occupancy prediction. Given the current submission, both aspects seem underdeveloped.

1. what is the contribution of BDC-V2 in Figure 3? It seems that it only increases the computational cost, with no performance improvement. Furthermore, MultiBiconv seems to be a multiple iteration of the technique from BDC-V1.
2. The paper seems to propose the design of a binarization methodology, not a design for Occupancy Prediction. While this is the first application of binarization to occupancy prediction, other binarization methodologies seem to be easily adaptable.
3. The design in Section 3.4 seems to be very similar to that of BiSRNet, which may limit the contribution of the paper. Is there anything different about the design of BiSRNet?

• In terms comparing the FPS with FlashOCC, one of the main objectives of binarization is to provide FPS speedups, since the provided results are marginally improved at best it seems to make the proposed approach less necessary. Instead FPS should be compared with other binarization methods as well for fairness.
• From the current writeup it is not clear how the proposed module can be plugged into other existing methods.
• I am unconvinced that since in my understanding the backbone is left as FP32 why binarization is necessary and other forms of quantization are not considered appropriate.
• It is not clear what aspects of the approach are specific to the occupancy task and what is generalizable to other tasks as well.

1. Limited exploration of binarization strategies: The paper primarily focuses on binarizing convolutional layers using the BiSR-Conv method. Exploring alternative binarization techniques, such as XNOR-Net [1] or DoReFa-Net [2], and comparing their performance with BDC-Occ would strengthen the analysis and potentially reveal further insights. It's also important to investigate the impact of different activation functions specifically designed for BNNs, like those proposed in ReactNet [3].

2. [minor] Lack of comparison with other compression techniques: The paper positions BDC-Occ as a solution for deploying occupancy networks on edge devices. However, it lacks a comparison with other model compression methods beyond binarization, such as pruning [4], quantization [5], or knowledge distillation [6]. Demonstrating the advantages of BDC-Occ over these alternatives would significantly bolster its practical significance.

3. [minor] Limited generalizability: The paper acknowledges the limitation to CNN architectures. However, this limitation needs further discussion and investigation. This would help understand the challenges and potentially open avenues for future research on binarizing more diverse network architectures.

References:

[1] Hubara et al., Binarized Neural Networks, NIPS 2016.

[2] Zhou et al., DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients, arXiv 2016.

[3] Liu et al., Reactnet: Towards precise binary neural network with generalized activation functions, ECCV 2020.

[4] Han et al., Learning both Weights and Connections for Efficient Neural Networks, NIPS 2015.

[5] Jacob et al., Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference, CVPR 2018.

[6] Hinton et al., Distilling the Knowledge in a Neural Network, arXiv 2015.