Summary and Contributions: In this paper, a novel collaborative aggregation-and-distribution network is presented to detect the co-salient objects from an image group. The main contributions are three-fold, which attend to several major issues in the task of co-salient object detection. First, an OIaSG is constructed to combine saliency detection with co-saliency detection in a unified joint-optimization framework. Second, the paper proposed a two-step aggregation-and-distribution strategy to mine more discriminative group-wise information by GASA and adaptively capture co-saliency cues by GGD. Third, the authors strengthen the group image relationships in the process of feature up-sampling. The experiments show competitive performances both qualitatively and quantitatively. After reading rebuttal information. ========================== The authors clearly addressed my comments. I have also read other reviewers’ comments as well as the corresponding responses. The additional experiments and analyses seem to be convincing. I think the paper is interesting, which may be of interest for the researchers in the related areas.
Strengths: This paper investigated the critical problems of co-saliency object detection, including the input order issue, inter-image modeling issue, intra and inter information integration issue, and group consistency issue. The claims made by the authors are intuitive and turn to be effective. A novel GASA module was constructed for capturing robust group semantics, which remains unchanged facing varying input orders. The design of GASA also followed multi-head attention and included non-local aggregations to learn more powerful group-wise features. The GGD and GCPD further provided new ways of designing better deep architecture for the co-saliency detection task in terms of fully modeling inter-image interactions across all stages. The ideas of this paper are novel and insightful, the flowchart is reasonable and reproducible, and the experiments are sufficient and competitive.
Weaknesses: For the CoSOD task, some related images in an image group are used as the inputs of the model, which is similar to the video SOD task. Could you directly transform the CoADNet or some parts to the video SOD task? The authors claim that the jointly-optimized OIaSG module could provide more reliable saliency priors. Traditionally, a typical practice would be to pretrain the backbone network on SOD tasks to gain useful priors. However, it seems that the authors did not analyze why the proposed OIaSG could be a better choice. In other words, its effectiveness is not fully evaluated. Some of the experimental setups are not detailed enough. For example, the network takes five images as a sub-group, which means that it is often impossible to process all images from a certain category. In such context, how the images are sampled and exploited should be clearly explained. In the experiments, only the inferencing time is provided. How about the model size of the proposed method compared to the SOTA model, such as GCAGC? Will you release the source code in the future?
Relation to Prior Work: yes
Summary and Contributions: This paper presents an end-to-end collaborative aggregation-and-distribution network (CoADNet) to capture both salient and repetitive visual patterns from multiple images. The proposed model contains an online intra-saliency guidance (OIaSG) module for supplying saliency prior knowledge, a two-stage aggregate-and-distribute architecture to learn group-wise correspondences and co-saliency features, and a group consistency preserving decoder (GCPD) to replace conventional up-sampling or deconvolution driven feature decoding structures. ---------after rebuttal------- 1) I have read the response and other reviewers' comments. While most minor concerns are addressed by authors, I still feel this paper is weak in high level insights and novelty. In my view, this work seems like an integrated version of most recent works, including group shuffling (simply applying channel shuffle  to multiple group of features) and group attention (multi-head attention). 2) Moreover, Table. 2 shows that with OIaSG, the model achieves significant performance boost (Line 285-288: 8.4% on Cosal2015 and 5.5% on CoSOD3k performance improvements according to F-measure). Yet, please also note that OIaSG requires additional labeled data for training, which are acquired from SOD dataset (the largest SOD dataset: DUTS). Therefore, it is not clear if the performance gains are from additional data and, if it is fair to other compared SOTAs.
Strengths: -The authors present comprehensive analysis about current issues of co-salient object detection. -The proposed method achieves state-of-the-art results on Cosal2015, CoSOD3k, MSRC and iCoseg datasets. -The paper is easy to read and follow.
Weaknesses: -The overall motivation of this work is not clearly stated after reading introduction. It is like an integrated model that exploits several incremental techniques to improve the performance. -The section of related work is missing. The authors should summarize the typical strategies about co-salient object detection, and some relevant modules in most recent works. -The experiments are not convincing. The presented results seem cannot support the claimed contributions, it's very hard to say what kind of feature is learned with the proposed method. The authors should provide some feature visualizations to verify the mentioned benefits. -There is a complete lack of discussing the impact of adding additional parameters and additional computational effort due to several modules appended on the baseline. The authors should provide this analysis for a fair comparison with other works. -In addition, this work is weakly relevant to the NeurIPS community.
Correctness: This paper is technically sound.
Clarity: This paper is generally well written.
Relation to Prior Work: Without related work. Not convincing.
Additional Feedback: after reading the rebuttal and reviews from other reviewers. I still feel the overall design is too straight-forward and does not seem to be very interesting/exciting. Therefore, I can not vote for acceptance,
Summary and Contributions: This paper proposes an end-to-end CoADNet framework for co-salient object detection (CoSOD), which produces state-of-the-art performance on four popular benchmark datasets. The key components of the proposed CoADNet include aggregating group-wise semantic representations and distributing the learned group semantics to individual image features. It uses the OIaSG, GASA, GCD, and GCPD modules to address the following issues: 1) how to bridge the relationship between saliency and co-saliency; 2) how to generate more discriminative group-wise features that are insensitive to the input order; 3) how to distribute group semantics to different individuals; 4) how to maintain inter-image constraints during feature decoding stage.
Strengths: +The authors discussed several important issues, which is not considered or well addressed in previous studies, in my opinion, is very inspiring for the CoSOD tasks. For example, an intuitive idea about group feature learning is that it must be order-invariant, and the GASA module elegantly solves it and also keeps efficient. +Another interesting idea is that the learned group representations should be adaptively distributed, instead of being directly duplicated for concatenation. +Previous CoSOD works do not consider whether it is necessary to consider co-saliency constraints during feature decoding, and this paper shows that it is also practical to further boost the performance by tailoring specific decoders for co-saliency cues mining.
Weaknesses: -The designed CoSOD framework only accepts fixed-number (i.e. 5 images) group-wise inputs at each time. How to handle the case during testing where the number of samples in a particular category is a non-integral multiple of 5? Is there any overlap between the data? -In the GASA module, there is a pre-defined parameter “b” which controls the number of partitioned feature blocks. However, the authors did not explain how to choose an appropriate value for “b” and to what extent will it influence the model. Besides, in the second paragraph of Sec 2.3, how to perform “block-wise group shuffling” is vague. More detailed descriptions are needed. -In the CoADNet framework, authors have repeatedly emphasized that the model is not sensitive to the input order, but has not specifically explained how it works. -The proposed CoADNet uses ResNet50 and Dilated ResNet50 as backbone networks. However, some of the competing methods only use less powerful backbone such as VGG-16. For fair comparison, the authors should also provide the experimental results with VGG-16 as backbone network.
Correctness: Most claims look correct to me.
Clarity: Yes, the paper is well written and easy to follow
Relation to Prior Work: The main difference between this work and the exisiting works is clearly discussed but some existing literatures on COSOD are mising.
Additional Feedback: ---------------------------update after rebuttal---------------------------- In the authors’ feedback, the authors addressed my concerns on technical details and experimental comparisons. They implemented experiments to use VGG-16 as their backbone and the obtained performances are better than other state-of-the-art methods. Overall, the novelty, technicality, and performance of this submission are good to me. So I would like to keep my initial score and favor accepting it.
Summary and Contributions: The authors proposed a network consisting of different variants to solve the task of Co-Salient object detection, which leads to improved performance. More specifically, the proposed pipeline integrates saliency priors, group-attentional semantic aggregation to achieve superior performance. Experiments on different datasets under various settings demonstrate the effectiveness of the proposed approach.
Strengths: 1. The combination of different components has shown to perform well to solve the task of co-saliency detection. 2. Joint optimization of intra-saliency guidance and the co-saliency query group helps to suppress the background information. 3. The paper is relatively well written and easy to follow. 4. A reasonable ablation study is provided to justify different variants used in the proposed method.
Weaknesses: 1. While the idea of integrating saliency priors, group-attentional semantic aggregation modules for Co-Saliency object detection is interesting, there is a complete lack of discussing the impact of adding additional parameters and computational effort due to the different variants used in the proposed complex pipeline. The authors should provide this analysis for a fair comparison with the closest baseline . 2. I was wondering why the numbers reported in Table 1 for  are different from the results reported in . Is it because  used VGG-16 as a backbone network while the authors used ResNet50 as the backbone. I am confused why the authors did not add a column to include the backbone network for each baseline method. For a fair comparison, I strongly suggest authors should experiment using the VGG-16 backbone to demonstrate superiority over the baselines. It is not clear if authors implemented baseline methods e.g.  using ResNet50 backbone. To show the superiority of the proposed method, I strongly believe that authors should report results using the VGG-16 backbone rather than implementing baselines using ResNet-50 backbone (if the authors used ResNet-50 to report baseline numbers). 3. More importantly, the idea of group semantics for co-saliency detection has been already explored in . 4. Additionally, I was wondering why the authors did not report results using the AP metric. Is there any specific reason? 5. The idea of incorporating saliency priors to suppress the background information is nothing new in the context of saliency detection, semantic segmentation.
Correctness: The authors should clearly mention what backbone network each baseline method use for a fair comparison.
Relation to Prior Work: Not strong enough.
Additional Feedback: While the idea of combining different components is interesting in the context of co-saliency detection, it is hard to judge whether the proposed approach indeed demonstrates state-of-the-art results without the help of a stronger backbone network. The method outperforms the baselines probably due to the ResNet-50 backbone while other baselines methods used VGG based backbones. The reported numbers in experiments suggest noticeable improvements; however, there are several major issues mentioned in the weakness section that need to be explained before the final rating. After reading rebuttal information. ========================== While I am a bit more positive about the paper after reading the rebuttal and other reviews, my original concern regarding the technical novelty and contributions persist. The proposed model exploits several incremental techniques (group shuffling and group attention) to improve the overall performance as agreed on by R2. So, I am not convinced the paper has enough novelty to be presented as a new contribution to the NeurIPS community.