Dear Reviewer XRVn:
We appreciate your detailed summary and positive comments regarding the presentation, novelty, creativity, and the experimental results of our research. Thanks a lot for providing the valuable feedback and raising insightful questions about our work. We address and answer your questions below:
**Q1: Could you provide the overall inference speed and the speed of each module? Additionally, how much does the Doppler bins descriptor reduce the overall computational load, and do the range-wise self-attention and spherical-to-Cartesian feature aggregation introduce significant computational overhead?**
A1: Thank you for raising this concern. Please first refer to our general answer to the **‘Computation complexity and requirement’** question in the global rebuttal.
Specifically, the range-wise self-attention module is highly efficient, with a runtime of only 2.5ms. The spherical-to-Cartesian feature aggregation initially costs 72ms per frame. However, this can be optimized to 29.7ms by reducing the number of layers. This module not only converts the spherical feature volume to a Cartesian one without interpolation errors but also enhances the encoding of voxel feature relationships.
The Doppler bins descriptor significantly reduces the data volume of raw 4D radar tensors (4DRTs) by a factor of D/8, where D equals 64 in our case. This reduction decreases the input data volume to 1/8 of its original size. This reduction not only shortens loading times (from disk/sensor to processors) but also lessens the overall computational burden during spherical-based feature encoding.
**Q2: Dependence on a Single Dataset. The experiment was limited because of the data set**
A2: Thanks for highlighting this concern. Please refer to our general answer to the **‘Dependence on a single dataset’** answer in the global rebuttal.
**Q3: Author just use well-condition sequences, which I think may be unfair to lidar-based methods or vision-based method**
A3: Thanks for this thoughtful feedback. We are not sure if we understand this question exactly. If you were concerned that the training data used to train baselines method does not include sequences collected from adverse weathers, which causes worse performance due to the domain change of LiDAR and camera data, we are happy to answer this question as follows.
Currently, it is hard to resolve this problem as the LiDAR data is inevitably affected by the adverse weather. For example, water droplets or snowflakes can scatter or absorb LiDAR beams, reducing the effective range of LiDAR and inducing noise in the data. As a result, the occupancy labels obtained from accumulated LiDAR data under adverse weather have low fidelity. The nuScenes dataset also faces a similar issue where occupancy labels are affected by adverse weather, despite having annotated bounding boxes. Adding new nuScenes data collected under adverse weather to train our LiDAR and vision-based baselines might not be feasible. This approach could potentially hinder network training because of noisy labels rather than improve adaptation to adverse weather. Moreover, as nuScenes does not provide 4D radar tensor data, it’s unfair for our radar-based methods because they cannot benefit from the new data samples.
This situation highlights a significant research question: how can we generate reliable 3D occupancy labels under adverse weather conditions when LiDAR measurements are noisy? We acknowledge the importance of this question and plan to address it in our future work.
**Q4: Could you provide the overlap ratio between different sensors and ROI?**
A4: Thanks for this comment. Our evaluation focuses on the overlap area between the horizontal field of view (FoV) of all sensors and our defined RoI to minimize potential data discrepancies beyond the FoV. Specifically, the overleap hFoV of K-Radar sensor suite is 107◦, symmetrically distributed around the front axis. The ratio of between the final evaluation area and our RoI can be calculated as:
$$ 1 - \frac{\cot\left(\frac{107^\circ}{2}\right)}{4} \approx 0.812 $$
which means 81.2% area of our RoI is taken into account for our final evaluation area.
**Q5: Any way to improve training data and generalization**
A5: Thanks for this insightful question. It aligns with our research interest in efficiently training 3D occupancy prediction models. Some recent works are exploring self-supervised methods in occupancy prediction. One notable example is EmerNeRF (https://emernerf.github.io/), which can represent highly dynamic scenes in a self-sufficient manner. This approach has the potential to label occupancy ground truth using RGB cameras without requiring human annotation efforts.