Summary and Contributions: The paper presents an approach for one-shot video object segmentation that replaces the pretraining step on large image segmentation datasets with a meta-task that learns an optimal network initialization for one-shot fine-tuning and other hyper-parameters such as learning rates. The claimed advantage is speed, evaluation on the DAVIS datasets demonstrates that the proposed approach has a better speed/accuracy tradeoff than other OSVOS approaches that employ finetuning but worse than the SOA network STM, that uses attention instead of finetuning.
Strengths: Novelty - to my understanding this is the first time that MAML inspired meta-learning is used for OSVOS. The results and the ablation study seems to demonstrate its effectiveness.
Weaknesses: While the paper’s major claim is efficiency, intended as inference time, there is no treatment of the topic neither in the paper nor in the supplemental material beside the plot of figure 1 (change power of ten to integer number in the x-axis notation). Why is the algorithm more efficient? Is the fine-tuning faster to converge because of the learned initialization? How many fine-tuning interactions are needed on pre-trained networks? How many with meta-learning? How accuracy changes increasing number of iterations with the proposed approach? Many of the methods that are compared in this paper strived for accuracy - I suspect their finetuning might have been much faster, reducing the number of iterations with negligible loss of accuracy. Also, is the runtime speed amortized over the number of frames? If so, the longer the video the less relevant is the time spent finetuning. Furthermore, an important baseline is missing. I would have liked to see the performance of the proposed approach, with all the bells and whistles and with pretraining on ImageNet or any other large scale image segmentation dataset instead of the proposed meta-learning. That’s crucial to assess the actual benefits of meta-learning compared to pretraining. Neuron-Level Learning Rates - to my understanding a learning rate is learned for each kernel and not for each neuron. If that’s the case, please improve the wording, which is confusing.
Correctness: The proposed method and evaluation is technically sound.
Clarity: The paper is well written.
Relation to Prior Work: The discussion of related works is comprehensive.
Summary and Contributions: The paper proposes three improvements to finetuning-based video object segmentation. The contributions include 1) a meta learning strategy for learning a better initialization weight, 2) the learning rate is predicted with learnable parameters, and 3) the online finetuning is done on the detection model directly. *after rebuttal*, I have increased my rating to 6, as I believe it's overall an interesting combination of VOS and meta-learning. The rebuttal is also very helpful, especially fig1. i encourage the authors to include the missing details if accepted.
Strengths: The techniques proposed in the paper are quite intriguing. Especially, the idea of using meta-learning to learn a better initialization is well-motivated, and seems to be helpful as demonstrated in the experiments. The paper is well-presented, and the discussion with prior work is very comprehensive.
Weaknesses: The main concern is that, the validations in the paper are mostly done empirically, and lack of more in-depth analysis. Therefore, I am a bit skeptical of the effectiveness of the proposed methods, I believe more experimental analysis should be provided to confirm the efficacy. In particular, it would be interesting to demonstrate the difference of the weights learned with and without meta-learning. It might be also useful to show curves of J&F -- number of iterations in the test, to see how the meta-learning can help the finetuning in the test time. For the learning rate predictor, it would be interesting to see in which circumstances it predicts a larger learning rate, etc. Some qualitative results should be provided as videos, to clearly demonstrate the improvements provided by the proposed components. Currently, due to that the analysis is largely missing, it is difficult to gain further understanding of the proposed methods. Thus I believe the insights from this paper is a bit limited.
Correctness: Seems correct to me.
Clarity: The paper is well-written.
Relation to Prior Work: The contributions are clearly summarized.
Summary and Contributions: The paper addresses the problem video object segmentation in the semi-supervised setting, meaning that the instance segmentation masks for the first frame are given during inference time. High performing methods in this field tend to fine-tune a model based on the first frame during inference, resulting in high latency. This paper proposes several tricks to reduce the fine-tuning overhead. One key idea is to meta learn the model initialization and fine-tuning learning rates for the test time optimization. In fact, every neuron in the network receives a separate learning rate. Moreover, the method fine-tunes on not only the first frame, but also subsequent frames using inferred masks. Experiments were done on DAVIS-2016/2017 + Youtube-VOS. These are all common benchmarks. The authors claim novelties in: - Learn the model initialization instead of just training an auxiliary model and start from a random checkpoint. - Learn the learning rates: a learning rate is learned for each neuron in the network. - It uses Mask R-CNN that first detect objects and infer the masks. This way, no global semantic segmentation is necessary. I agree with the authors' assessment regarding the contributions.
Strengths: - Problem statement: video object segmentation is a well researched problem in the computer vision community, and is therefore relevant to NeurIPS. - Novelty: AFAIK, I do not know any previous work that attempts to optimize for the model initialization and learning rates that are used to fine-tune for the first frame. I believe that this approach is indeed novel. The meta learning framework seems sound as well. - Sufficient experimentation: the paper demonstrated strong results on 3 benchmark datasets, and it has a good ablation in table 1 as well. - Strong results: the method outperforms other methods (with the exception of STM) by a large margin on all 3 datasets.
Weaknesses: There are some clarity related issues in the method section. The goal of the method is to learn the model initialization and the learning rates for the neurons. One typically cannot learn the learning rate. The reason why is because there is no trivial gradient flowing from the loss to the it. In this paper, the learning rates are learned. So why is there a gradient wrt the learning rate in the first place? I am positive that the authors have somehow solved the problem, but I have a hard time finding it in the paper. It's a shame, but without fully understanding how the gradients wrt to the meta parameters are derived, I cannot give a better score than borderline.
Correctness: Mostly correct, except for the clarity issue mentioned above.
Clarity: Some clarity issues as mentioned above. Otherwise, good presentation.
Relation to Prior Work: The paper gave a thorough outline of existing methods for video object detection, including methods with or without finetuning.
Additional Feedback: ln 7: To mitigate what? The previous sentence says that most methods refrain from doing test time optimization. So mitigating refraining from optimization? Table 1 says fine-tune epochs? Since we are doing single-shot training, is one epoch equal to one image also? In ln 167, does g refer to the learning rate and number of steps? The benchmark datasets are all fairly small. Did the authors observe any overfitting when doing meta learning? **************************************************************************** POST REBUTTAL / REVIEWER DISCUSSION COMMENTS Thank you for providing the rebuttal. I am satisfied with the rebuttal, but I also concur with my fellow reviewers, R1 and R2, on the missing detailed latency analysis and ablation. I'll improve my rating slightly. Best of luck!
Summary and Contributions: In this paper, an efficient approach of online learning based video object segmentation is proposed. The proposed method starts from OSVOS , and make some essential changes to make it efficient. The major changes are changing the backbone and adopting meta learning for initialization. The proposed method show competitive performance to the state-of-the-art methods on VOS benchmarks.
Strengths: The proposed method is based on online learning but efficient. Obviously, the recent trend in VOS is offline approaches that do not rely on online learning. it is because online learning has an important shortcoming that slows down the inference. In this paper, instead of following the recent trend that adopts the offline method, the authors approached to solve the disadvantages of the online method through meta learning. I want to admit that part. it is because the online learning based method is better than the offline learning method in some aspects. For example, it generalize to unseen objects better than offline methods as shown in Table 2, J unseen.
Weaknesses: The performance, in terms of both the accuracy and the speed, lags behind the recent state-of-the-art offline methods.
Correctness: The proposed methodology look correct.
Clarity: The paper is written clearly.
Relation to Prior Work: Yes.
Additional Feedback: The implementation looks very challenging. The author is strongly encouraged to publish their source code.