The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation

Denoising diffusion probabilistic models have transformed image generation with their impressive fidelity and diversity. We show that they also excel in estimating optical flow and monocular depth, surprisingly, without task-specific architectures and loss functions that are predominant for these tasks. Compared to the point estimates of conventional regression-based methods, diffusion models also enable Monte Carlo inference, e.g., capturing uncertainty and ambiguity in flow and depth. With self-supervised pre-training, the combined use of synthetic and real data for supervised training, and technical innovations (infilling and step-unrolled denoising diffusion training) to handle noisy-incomplete training data, and a simple form of coarse-to-fine refinement, one can train state-of-the-art diffusion models for depth and optical flow estimation. Extensive experiments focus on quantitative performance against benchmarks, ablations, and the model's ability to capture uncertainty and multimodality, and impute missing values. Our model, DDVM (Denoising Diffusion Vision Model), obtains a state-of-the-art relative depth error of 0.074 on the indoor NYU benchmark and an Fl-all outlier rate of 3.26\% on the KITTI optical flow benchmark, about 25\% better than the best published method. For an overview see https://diffusion-vision.github.io.

Paper

References (88)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer LqWs7/10 · confidence 4/52023-06-29

Summary

This paper approaches the task of predicting optical flow from a pair of images and depth from a single image. It proposes to do so using diffusion models, providing a training pipeline which includes contributions to deal with noisy training data. The proposed method is competitive with SOTA in depth prediction on NYU and KITTI, is similar or better to SOTA in optical flow prediction on Sintel and KITTI zero-shot, and is SOTA in optical flow after finetuning on KITTI. Additional experiments show the proposed pretraining pipeline meaningfully improves RAFT and the proposed model, that contributions dealing with noisy training data are helpful, and that the proposed method can predict multimodal outputs in cases of uncertainty. Edit: Thanks to the authors for the rebuttal. After reading the rebuttal and other reviews I will keep my rating at 7 - accept. I believe this paper should be accepted because it (1) introduces a simple but interesting method that performs SOTA on competitive benchmarks and (2) is very polished and experiments clearly defend all main contributions of the method.

Strengths

The paper provides a simple method that is similar-to or better than SOTA on competitive optical flow and depth tasks - A new data pipeline is introduced which significantly boosts performance of prior SOTA e.g. RAFT, while improving further with the proposed method - The method is built off a standard image-to-image diffusion method Palette, adds minimal detail needed to handle flow and depth data (e.g. L1 loss, infilling + unrolling, coarse-to-fine), and is competitive across a wide variety of experiments - Real data is tricky to handle for diffusion models as it can contain incomplete flow and depth maps. The proposed method has an interesting idea of using its own predictions late in training. Combined with infilling, this is helpful. The paper is very polished and experiments clearly defend all main contributions of the method. - Writing is very clear, figures and tables are attractive and helpful - Choice of training data is very important (Figure 3, Table 6, 7); it even improves RAFT meaningfully (Table 1) - The method can also produce Multimodal samples faced with ambiguity e.g. transparent/translucent/reflective (Figure 1, 4, 5) - Figure 6 and Table 5 shows the importance of coarse-to-fine refinement, which enables the method to produce better detail than RAFT (Figure 4, 5) - Infilling + step-unrolling yields massive improvement on KITTI (Table 4) - L1 loss (Supp Table 4), pretraining for depth (Supp Table 5, 6) - Zero-shot depth completion is a cool application of the method

Weaknesses

Finetuning performance on real data yields weaker performance on Sintel vs. other methods (minor weakness, but addressing could further strengthen paper!) - On Sintel (Table 2), most other methods use warm start, i.e. initializing flow prediction from previous frame. Is it possible to do this with diffusion, i.e. starting from a previous prediction, perhaps denoising fewer steps? If not, this is an important weakness of diffusion models in this setting - Given most other method warm-start, it is harder to analyze the ability of the proposed method to finetune on real data. Perhaps it does not do as well relative to FlowFormer because (1) it does not handle real data as well as regression methods, even after contributions in this area, or (2) it does not overfit as well to specific datasets given others tend to have task-specific architecture. A comparison to more methods without warm-start, or by using the proposed method with warm-start, it would be very interesting to see this analyzed. The main novel contribution of this paper beyond data pipeline is step-unrolled denoising diffusion training, which yields modest improvement over infilling alone (minor weakness) - L1 loss is a design decision, infilling holes using bilinear interpolation is not a substantial contribution in my eyes. So the remaining technical contribution is unrolling. - I understand the contributions are deliberately simple, which is a positive given performance gain. However, it is still important to analyze the reasons for the method’s success. In this case, the most novel component, “unroll”, improves optical depth REL from 0.077 to 0.075, RMS from 0.338 to 0.324, KITTI REL from 0.057 to 0.056, RMS from 2.744 to 2.700, AEPE from 1.53 to 1.47, and F1-all from 5.24% 4.74%. These contributions are nontrivial, but not substantial on their own.

Questions

- L179 is a bit confusing and could be rewritten “Training high resolution diffusion models is often slow and memory intensive but model performance has been shown to improve with resolution” - Figure 4 (top) is dark and indistinct, making it a hard to see. Are there failure cases for RAFT on more easily readable examples?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Yes

Reviewer F9qi7/10 · confidence 5/52023-07-05

Summary

This paper proposes to use diffusion models to solve monocular depth and optical flow estimation tasks. Unlike previous task-specific models for depth and flow, this paper uses a generic diffusion model. This paper studies the effect of training data (synthetic and real) and processing of sparse depth and flow ground truth when training the diffusions models. Experiments are conducted for depth and flow tasks on representative benchmarks, the proposed method achieves state-of-the-art depth performance on NYU and state-of-the-art optical flow performance on KITTI.

Strengths

- **The idea of using diffusion models to solve depth and optical flow tasks is interesting.** Depth and optical flow are typically approached as regression tasks, it's unclear how the popular diffusion models will perform for both tasks. This paper explores this direction and shows some interesting results. - **Several strategies are proposed to handle the issue of data for training diffusion models.** It's not straightforward to apply diffusion models to depth and flow tasks, the paper proposes data infilling, step-unrolling and an L1 loss to tackle the challenges. - **The experiments are extensive and informative.** Training data plays a significant role in training diffusion models, this paper studies the effect of different training datasets for both depth and flow tasks. The performance on KITTI for optical flow task is especially strong, outperforming previous 2-frame optical flow methods by a large margin. - **A detailed discussion of limitations is presented in the supplementary material.** This paper gives a deep analysis of the limited performance on Sintel test set and the results indicate that a particular sequence on the test set severely affects the averaged performance, which might provide some hints for further improvement in future.

Weaknesses

I didn't observe major weakness and would put some minor points to the Questions.

Questions

- Due to the uncertainty in diffusion models, a same model might produce different results when running twice. How large is the fluctuation and how's the final quantitative results reported? Are the authors using some averaging? - How sensitive of the proposed architecture to different image resolutions for both depth and flow tasks? For example, what if the inference image resolution is different from training, will the model still perform reasonably? - The optical flow results on KITTI is very strong, but the results on Sintel seem less robust (as also analysed in the supplementary material). What might cause the different behaviours on Sintel and KITTI, could the authors further comment on this? - I think one key message from this paper is that the experiments show the importance of training data. When comparing Table 1 and Table 6 for the results of RAFT and the proposed method, we can observe that RAFT outperforms diffusion models when only using AutoFlow for pre-training. However, diffusion models perform better when more datasets are added to the pre-training stage. I am wondering whether this indicates that diffusion models can benefit more from larger datasets than previous regression methods?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Yes, the authors have carefully analyzed the limitations.

Reviewer xrrD8/10 · confidence 3/52023-07-08

Summary

The authors study the use of diffusion models for the tasks of single-image depth estimation and optical flow estimation. Self-supervised pre-training, supervised fine-tuning with synthetic and real data, combined with a couple of tricks to leverage imperfect GT, lead to competitive results with nearly no task-specific modifications to the diffusion models. The authors also demonstrated unique capabilities enabled by the diffusion models, e.g., capturing multimodality and completion from partial data.

Strengths

**Originality and significance**: As far as I know, this is the first work to study the use of generative diffusion models for optical flow and depth estimation tasks. Training good optical flow and depth models typically requires extensive task-specific knowledge in terms of architecture designs and loss functions, so the authors' finding that diffusion models can compete well on these tasks with almost no task-specific treatments is non-trivial and valuable. The competitive results in various settings further add to the significance of the work. **Technical quality**: Training diffusion models successfully involves many technical details. The authors generally follow best practices and propose reasonable solutions to unique challenges. More specifically, the use of pre-trained PALETTE, further supervised pre-training with mixtures of synthetic and real datasets, addressing imperfect GT with infilling and step-unrolling, etc. are all well-motivated and proven effective. **Writing quality**: The paper is nicely written, with precise language, adequate details, and clear explanations. Conclusions are justified with plenty of results, visualizations, ablation studies, and overall convincing.

Weaknesses

The authors claim that diffusion models can be a generic framework for vision dense prediction tasks, but only consider the tasks of depth estimation and optical flow in this work. Both these tasks are relatively low-level, and it'd be interesting to offer some insights, discussions, or analyses regarding how higher-level tasks, such as semantic segmentation, differ from them and if the claim still holds. Two relatively minor complaints/suggestions: * Can authors provide some theoretical analysis or proof for the step-unrolling step? Could unrolling more steps bring further improvements? * It's unclear how much the use of pallete self-supervised pre-training help since it's not part of the ablation study.

Questions

Please refer to the three points in the weakness section above.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

I agree with the limitation the authors brought up in the supplementary. Optical flow and depth estimation are low-level tasks commonly used at early stages of real-world application systems and therefore demand higher efficiency. As the authors already pointed out, this is where the proposed diffusion models fall short.

Reviewer 8jCd6/10 · confidence 3/52023-07-09

Summary

The paper demonstrates that diffusion models are effective general-purpose solutions for dense optical flow and monocular depth regression tasks. The paper shows that the same architecture and loss functions lead to at-par or better performance on these tasks, compared to existing methods that use domain knowledge and problem-specific architectures. The main insights presented are the use of a pre-training phase for higher quality, and imputation of the missing values, and a step-unrolled diffusion step, for dealing with incomplete GT training data.

Strengths

The paper is well-written and motivates the contributions perfectly. The simplicity of the architecture and the loss functions are very clear. The results support the claims and show at-par or better results than the state of the art. Extensive experiments on different real datasets help evaluate the method's quality. Multi-modality of the outputs is promising! I would have loved to see some more analysis there. The presented application is also exciting, showing directions for text to 3D reconstructions.

Weaknesses

While there is limited technical novelty, this is a good paper that demonstrates a simple method for solving two different regression tasks. - I did not understand why the pre-training phase needs to be separate. The method uses supervised pre-training, that is different for each task. The loss functions are also identical between the pre-training and the fine-tuning states. Why not combine all available datasets and just train the model once (inc. all the tricks used for fine-tuning)? I did not understand this distinction between phases, especially when the experiments are explained and the datasets keep moving from one phase to the other. - Step-unrolled diffusion is a little similar to self-guidance, introduced in "Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning" [Chen et al.] I wonder if self-guidance is already good at addressing the domain shift problem, or whether step-unrolled diffusion is really needed.

Questions

- Please answer the questions raised on pre-training, and on self-guidance. - Is it possible to fine-tune an existing diffusion model, such as stable diffusion, instead of training one from scratch? Could it help avoid the use of synthetic data?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer F9qi2023-08-11

Thanks for the detailed responses. Overall I agree with the authors, but would like to have a followup discussion on one small point. Regrading the limited performance on the Sintel ambush_1 sequence, the authors mentioned that "FlowFormer’s inductive bias (with the cost volume etc) make it better suited to reason about this sequence’s large out-of-frame motions". However, I guess the cost volume will also be less effective for out-of-frame motions since the points are not matchable across frames? Thus it seems less likely that encoding the matching cost in the architecture will solve this problem? Maybe leveraging some context information (e.g., motion smoothness/propagation in GMA and GMFlow) would be helpful? I think this is an open question and could be considered as future work. This question doesn't affect my opinion, I am happy to accept this paper. Thanks.

Authorsrebuttal2023-08-14

Thank you for your comment. We agree with the reviewer that motion propagation and global motion aggregation are likely the main signals needed. We provide more details below. Broadly, we see two categories of errors: (1) From out-of-frame motion. As the reviewer suggests, the motion smoothness/propagation priors used in existing works might help with this. (2) Inconsistent flow for texture-less background regions. For example, ambush_1 frame_1 has a foreground that splits the background into two, with texture on one section of the background and almost no texture on the other. This leads to an ambiguity with our model sometimes predicting consistent flow for the entire background but often not (note that for frames with less ambiguity, the diffusion model is able to successfully propagate the motion). We find that the RAFT model also struggles on such sequences, which suggests that smoothness is not a sufficient prior for handling such cases. As the reviewer mentioned, GMA and GMFlow are some of the first models to successfully address this example (and furthermore the ambush_1 sequence) and it seems likely that this is due to their modules for global motion aggregation and motion propagation, since these signals can align the flow in the background. Theoretically, the self attention blocks in our model should be capable of global motion aggregation so it is possible that our training data distribution does not sufficiently cover such scenes. We hypothesize that there may be multiple ways to solve this problem (data, model design, improvements to sampling, i.e., better approaches to aggregation or re-ranking) and are excited to explore this in future work. The reason that the parenthetical in our original response starts with cost volume is because in current models the cost volume is the starting point for global motion aggregation modules (for FlowFormer, the aggregation happens as a post processing of the cost volume into a "cost memory", for GMA, the cost volume is processed into the "motion feature encoder" and then again into the "global motion aggregation" module), but the reviewer is correct to point out that the exact mechanism is global motion aggregation / motion propagation, which should have been mentioned.

Reviewer LqWs2023-08-15

Reviewer Response to Rebuttal

Thanks to the authors for the rebuttal. After reading the rebuttal and other reviews I will keep my rating at 7 - accept. I believe this paper should be accepted because it (1) introduces a simple but interesting method that performs SOTA on competitive benchmarks and (2) is very polished and experiments clearly defend all main contributions of the method.

Reviewer xrrD2023-08-20

Thank you for the detailed responses. It is a very interesting work with valuable insights and practical values. I don't see any major weaknesses in this submission. I maintain my acceptance recommendation and have no further questions.

Reviewer 8jCd2023-08-21

Thank you for the very well-written rebuttal. It clarifies the contributions and the other questions about the baselines and training details.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC