Response from Authors
Dear Reviewer,
Thank you very much for your valuable and prompt feedback. We sincerely appreciate the time and effort you have invested in reviewing our work and highlighting areas that require further clarification. We are eager to address your concerns as follows:
**Q1: Relationship between image size and memory usage.**
A: With PyTorch 2.2 and flash-attention, using 12 refinements and 4 frames, the GPU memory usage for StreamFlow is as follows. Specifically, when the image size increases by 4 times, the GPU memory usage increases to nearly 4.4 times. When the image size is increased by 9 times, GPU memory usage grows to nearly 20 times. For most current scenarios, StreamFlow still maintains a relatively moderate use of GPU memory. The variation in GPU memory usage may be influenced by underlying optimizations in the framework. We believe the memory usage could be further optimized in the future.
| Image Size | GPU Memory |
| ------------ | ------------ |
| 360x640 | 1.19 G |
| 720x1280 | 5.20 G |
| 1080x1920 | 24.11 G|
**Q2: All evaluation parameters for MPI-Sintel.**
A: We have collected all measures including ``all/matched/unmatched EPE, d0-10, d10-60, d60-140, s0-10, s10-40, s40+`` in Table 2 of the attched PDF. Compared to its two-frame baseline, StreamFlow performs exceptionally well in unmatched areas, validating its effectiveness in addressing occlusion issues. We would add a more detailed analysis for each metric in the supplements.
**Q3: Relative accuracy of StreamFlow.**
A: Thank you for your insightful comments and for pointing out concerns about the descriptions in this work. Our initial intention is to highlight its notable 0-shot cross-dataset generalization results, and we hope it did not come across as overstating its performance. We will refine the expression in the revised version to minimize potential misunderstandings. For instance, we will revise the content between Lines #223 and #230 to: ``From Table 1, we could learn that StreamFlow achieves excellent 0-shot cross-dataset performance on Sintel and KITTI. Compared to previous methods, StreamFlow reduces the 0-shot end-point error by 0.16 and 0.08 on the challenging Sintel clean and final passes, respectively. On KITTI, StreamFlow surpasses the previous 0-shot results with 0.11 and 17.65% lower EPE and Fl-all metrics. Besides, without self-supervised pre-training or bi-directional flows, StreamFlow attains commendable accuracy and efficiency on the challenging Sintel and KITTI test benchmarks using (C)+T+S+K+H schedule.``
Besides, StreamFlow indeed uses experimental settings different from others. However, this may also highlight StreamFlow’s advantages. For instance: (1) On the Spring dataset, StreamFlow was only trained for 180k iterations while outperforming MemFlow, which was trained for 400k iterations with the same batch size and learning rate. (2) For Sintel testing, StreamFlow was trained for 300k iterations on FlyingThings and 180k on T+S+H+K, while MemFlow was initially trained for 120k (FlyingChairs) and 150k (FlyingThings) in a 2-frame setting, followed by 600k iterations on FlyingThings and 600k on T+S+H+K. Despite this, StreamFlow still delivers good results, especially on Spring. (3) As for VideoFlow, it employs more frames and bidirectional flows for training, achieving excellent results on the KITTI dataset. We will update the result in Table 1 and add the related discussion. However, VideoFlow explores the accuracy and efficiency under bidirectional flow estimation, which differs from the focus of StreamFlow on a non-overlapping, continuous unidirectional flow pipeline. Additionally, its latency is significantly higher than that of StreamFlow. Therefore, A more relevant comparison for highlighting the issues that StreamFlow addresses is with the baseline method, Twins-SKFlow.
In the end, we greatly appreciate your valuable comments and the opportunity to clarify these points. Please kindly let us know if you have any further questions or require additional clarification. We highly value your insights and stand ready to provide any further information that could be helpful.