Siamese Masked Autoencoders

Establishing correspondence between images or scenes is a significant challenge in computer vision, especially given occlusions, viewpoint changes, and varying object appearances. In this paper, we present Siamese Masked Autoencoders (SiamMAE), a simple extension of Masked Autoencoders (MAE) for learning visual correspondence from videos. SiamMAE operates on pairs of randomly sampled video frames and asymmetrically masks them. These frames are processed independently by an encoder network, and a decoder composed of a sequence of cross-attention layers is tasked with predicting the missing patches in the future frame. By masking a large fraction ($95\%$) of patches in the future frame while leaving the past frame unchanged, SiamMAE encourages the network to focus on object motion and learn object-centric representations. Despite its conceptual simplicity, features learned via SiamMAE outperform state-of-the-art self-supervised methods on video object segmentation, pose keypoint propagation, and semantic part propagation tasks. SiamMAE achieves competitive results without relying on data augmentation, handcrafted tracking-based pretext tasks, or other techniques to prevent representational collapse.

Paper

Similar papers

Peer review

Reviewer 5RkZ7/10 · confidence 4/52023-06-27

Summary

In this paper, the authors propose a simple extension to Masked Autoencoders (MAE) to be able to pre-train on videos: SiamMAE. Two frames are sampled, independently encoded, and then asymmetrically masked. A transformer decoder is used to predict the missing patches in the masked image. The authors show that by masking a high proportion of patches (0.95) in the future frame and leaving the past frame unmasked, they are able to encourage the network to learn a more object-centric representation and focus on object motion rather than low-level image details. The authors show that this simple approach outperforms previous methods on many down-stream tasks and perform an extensive ablation to examine the architecture choices.

Strengths

Very well written paper. Some results are particularly impressive (e.g. more than 20% gain over VideoMAE) Simple but effective method and the honesty of presenting this as-is (instead of disguising the method as being more complicated than it needs to be) I think should be appreciated. Good comparison to other work and a great ablation section explaining many design choices

Weaknesses

Can only see two minor weaknesses: 1. The approach adds little over MAE and feels incremental, however it works very well and hasn’t been done before. 2. This approach relies on the temporal smoothness that is found in many curated datasets, such as Kinetics. However, for in-the-wild videos, with many sharp scene changes, this assumption becomes less likely to hold. However, it should be possible to split long videos into scenes in an unsupervised manner and then sample frames within a scene. It would be nice to a discussion on how this can be applied to large-scale internet datasets (which are becoming very popular for foundation models)

Questions

With standard MAE, it is possible to also mask the input images for the downstream tasks to speed-up computation. It seems that this avenue for speed-up would not be possible with this current approach since the previous frame has to be unmasked? Could the authors clarify the reason for not including temporal positional encoding (was this ablated?) It would be interesting to see a bit more how this frame-based method compares to video-based methods on downstream tasks (for example more comparisons like VideoMAE). Perhaps focusing on compute efficiency?

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 uZFR6/10 · confidence 4/52023-06-27

Summary

The paper proposes to use Siamese Masked Encoders for establishing correspondence for video input data. Uses the concept of predictive learning based on Masked Auto Encoder. Paper proposes to use asymmetric masking for present and future frames. Achieves best results in self-supervised setting for video label propagation tasks.

Strengths

Strengths: 1) The paper is well written for most of the parts and aptly elucidates the advantage of using a masked auto-encoder-based method for object-based correspondence. 2) Discusses in detail the architecture design choices for the encoder and decoder, ans uses a final design which is intuitive and simple and also focuses on relevant ablation studies. 3) The paper shows and achieves better results for object segmentation, part segmentation, and pose propagation tasks.

Weaknesses

Weakness: 1) Frame sampling: Table 3b discusses the effect of frame gap, however, it misses to comment on what could be the minimum frame overlap between the pair of frames for which predictions are out to be done. Apart from asymmetric masking, the degree of overlap between consecutive frames is also an important factor. 2) Fails to discuss the probable failure cases given the limitations from more qualitative results analysis belonging to various tasks.

Questions

Questions: 1) Ablation experiment: Comparison with FrameMAE. In case of FrameMAE, When we do provide 2 frames to the joint encoder, a better comparison would be comparing the asymmetric masking with FrameMAE, since from Table 2 (b, d) results as presented does not seem too behind for FrameMAE. 3) Section 4.4: This subsection would be more aptly titled as "Futher Insights" rather Qualitative Results.

Rating

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Overall the idea presented in the paper is lucid and necessary details regarding reproducing the experiments are mentioned. I think the paper puts a right step in the direction of improving well know correspondence (object level) problem using predictive learning. However I feel what this paper lacks is drawing a complete comparison for an image overlap percentage based study. Presenting zero-shot results for occluded scenes and object parts and discussing failure cases (as seen from qualitative results obtained)

Reviewer uZFR2023-08-17

Reply to Authors

Most of my concerns has been answered. I would expect authors to incorporate some of the changes pointed out (and accepted) by authors to be reflected in the final version of the paper. I would like to maintain my ratings.

Reviewer h4WK7/10 · confidence 5/52023-07-05

Summary

* This paper propose Siamese Masked Autoencoders for learning visual correspondence from videos called SiamMAE. * SiamMAE randomly sample a pair of video frames and randomly mask 95% of patches of the future frame, and the pair of video frames are passed into visual encoder(VIT), and cross attention decoder to reconstruct the target. * The authors conduct several experiments on downstream tasks(vos, human pose propagation), showing superior performance.

Strengths

* The paper presents a simple, yet highly effective method for the challenging video self-supervised framework. * The proposed method is well motivated and intuitive with excellent performance. * The ablation is sufficient and the writing is excellent.

Weaknesses

* Experiments on video recognition experiments should be reported. * Although SiamMAE is pretraining on video frames, i still think the authors should conduct experiments on image downstream tasks(such coco detection, segmentation) to show image representation capacity. * In Table1, The previous video ssl method(VPS) is also pretrained on Kinetics with ResNet50,however,the performance of SiamMAE with ViT-S/8 is similar in DAVIS & VIP & JHMDB, any explanation for this?

Questions

please refer to the strength and weakness

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

4 excellent

Contribution

3 good

Limitations

no limitations

Reviewer 9jyT7/10 · confidence 3/52023-07-06

Summary

This paper focuses on the self-supervised learning for video representations. The proposed SiamMAE operates on pairs of randomly sampled video frames and asymmetrically masks them, and then predicts the missing patches for visual representation learning. SiamMAE achieves significant performance and outperforms state-of-the-art self-supervised methods on video object segmentation, pose keypoint propagation, and semantic part propagation tasks.

Strengths

1 The motivation is clear and strong. 2 The proposed asymmetric masking and cross-self decoder are effective and achieve good performance.

Weaknesses

1 This paper mainly evaluates the proposed method on tracking problems. How is the performance on video classification tasks, such as UCF101 and HMDB51? 2 In Table 2(c), grid mask achieves better performance than random mask with 0.5 mask ratio. Increasing the mask ratio will improve the performance of random masking. How is the performance when increasing the mask ratio for grid masking? 3 This paper only investigates the ViT-S backbone. It is better to also leverage larger models, such as ViT-B, to verify the effectiveness of the proposed method.

Questions

My concerns mainly lie in the experiments.

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

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

3 good

Limitations

Yes

Reviewer e5XC7/10 · confidence 4/52023-07-12

Summary

This paper considers the problem of using self-supervised learning from video to obtain a representation that is well-suited to the task of estimating correspondence between a pair of images. They propose a significant modification of the MAE training procedure which is adapted for estimating correspondence: one image is not masked at all while the other has most (90%+) of its tokens masked, and the same (i.e. a siamese) encoder is applied to both. This is designed to require the model to internally establish correspondence. The model is trained for pixel prediction on Kinetics-400 and evaluated on several propagation tasks (object mask, part masks, human pose) using kNN inference to establish a dense correspondence field. SiamMAE is shown to greatly outperform existing self-supervised learning procedures with comparable backbone architectures, including methods trained on video. Ablative experiments confirm the importance of combining a siamese encoder with asymmetric masking. Visualization of the attention maps show that the model pays strong attention to object boundaries, seemingly a novel attribute.

Strengths

1. Good motivation and contextualization with respect to past work. 1. The modification of the MAE procedure is simple but clever, and manages to extract much more information for correspondence from video than past methods. 1. Visualization of predicted images is quite impressive (Figure 2), despite the main goal being to learn a feature extractor for correspondence. 1. Comprehensive evaluation with 3 different tasks and wide selection of relevant baselines. 1. Ablative experiments verify the importance of each component of the design. 1. Hyper-parameters provided for reproducibility.

Weaknesses

1. I'm not sure about the emphasis on predicting the _future_. It seems that the temporal order could be reversed (i.e. predict the past given the future) or randomized and I would expect similar results. Has this already been tested? 1. It wasn't abundantly clear how the patch-patch similarity was obtained. It seems to be taken from the cross-attention values within the decoder (line 239). However, this could be more clear since the decoder may contain multiple cross-attention layers with each having multiple heads? 1. It wasn't clear how k-NN and the queue were used to perform propagation. This should be explained in more detail or a reference provided. 1. No code provided at this stage. 1. Lack of confidence intervals (not a major issue - delta is quite large in most cases). 1. No evaluation of ViT-B and ViT-L (not a major issue - impressive results obtained with smaller model).

Questions

(Please also address or correct weaknesses above.) 1. What is the purpose of the [CLS] tokens? Is the model much less effective without it? (It is surprising that the attention masks for these tokens were so salient given that they are not involved directly in the loss.) Suggestions: 1. I wonder whether it's possible to identify a general principle/strategy for preventing shortcut learning, of which this is just one instance? Minor edits: (no need to respond) 1. Bold values in Tables 2d and 3b seem incorrect (e.g. 58.4 > 58.1, 56.7 > 56.5).

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

Additional limitation: 1. Even though Kinetics-400 was used without labels, it is possible that its image distribution is quite similar to that of the downstream tasks. It would be good to discuss this potentialityu, and say that the impact of dataset similarity has not been investigated? I do not see any negative societal impacts stemming from this individual paper.

Reviewer 9jyT2023-08-17

Most of my concerns are addressed. However, I still feel it is important to demonstrate the effectiveness of the method in the video classification task. HMDB51 is a small-scale dataset with only 3.5k training clips. Evaluating the method on this dataset will not introduce too much computational cost but can further verify the results.

Authorsrebuttal2023-08-21

Response to Reviewer

Thank you for your comments and suggestions. We hope to include some results on activity recognition benchmarks in the final version.

Reviewer 5RkZ2023-08-18

Dear authors, thank you for addressing the comments in my review. I feel you have adequately done so.

Reviewer h4WK2023-08-20

Thanks for the response, I have no further questions and will keep my initial rating.

Reviewer e5XC2023-08-21

> Confidence intervals I was referring to CIs for the distribution induced by the random variables in the training procedure (shuffling of training set, initialisation of model parameters), not inference. It would be best if these could be obtained from at least 3 trials for the final version. > [CLS] token Thank you for running the experiment without a [CLS] token. I understand that the [CLS] token may be useful in intermediate layers as a kind of position-agnostic "place" to accumulate global information, but it's a little unclear to me why the attention maps for the final layer would be meaningful if no loss is applied to the [CLS] token (Section 4.4). Is it possible to shed some light on this? Perhaps because the other tokens "use" the global information without necessarily "adding to it" in the final layer? This reminds me - are the query and key projections symmetric (i.e. $W_Q = W_K$) in the self-attention layers? > Impact of dataset similarity not investigated Thank you for acknowledging this. > Larger models Thank you for running initial experiments. > Future vs past Thank you for running this experiment; I hope it can be included in the final version. **Overall** The paper is intuitive, well-motivated and well-written. The authors have used the rebuttal to strengthen the paper. I do not have any major concerns and keep my initial positive rating.

Authorsrebuttal2023-08-21

Response to Reviewer Comment

Thank you for your comments and suggestions. We address the reviewer concerns below: >Confidence intervals Ideally, we would like to run all our experiments/ablations multiple times. However, we note that this is not a common practice in representation learning literature due high compute requirements per experiment. Nevertheless, we will re-run our base ablation setting 3 times to quantify the variance (if any) and include it in the revised version. >[CLS] token We don’t have a good understanding of why the attention maps for the [CLS] token are meaningful. Exploring this is an interesting research question which we leave to future work. >Are the query and key projections symmetric? No, we don’t use the same query and key weights.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC