Learning Fine-Grained Representations through Textual Token Disentanglement in Composed Video Retrieval

1. The comparative analysis with other methods is insufficient. In Section 5.2, Table 2 is presented but not mentioned or discussed in the text. This oversight limits the effectiveness of the comparison and fails to guide readers through the results.
2. Are all the experiments conducted solely on the introduced FineCVR-1M dataset? If so, this raises concerns about the generalizability and robustness of the proposed method. Relying on a single dataset may limit the findings and their applicability to other contexts.
3. Although the methods proposed by the authors demonstrate superiority over WebVid and EgoCVR in many aspects, it is essential to conduct experiments on existing datasets to further validate the effectiveness of FDCA. Testing on established benchmarks is crucial for establishing credibility and ensuring that the proposed method generalizes well to different scenarios.
4. Additionally, the results presented in Table A4 indicate that the authors' proposed method performs worse than CoVR in certain metrics. This raises questions about the robustness of the proposed approach.
5. To facilitate the authors' reading experience, it is recommended that all images be provided as vector graphics whenever possible. This approach will help prevent distortion when magnifying figures, ensuring that details remain clear and crisp. Specific figures that would benefit from this adjustment include Figures A10, A12, A14, and A15. Additionally, it has been noted that some of the fonts in Figures 2 and 3 are vector-based while others are scalar. This inconsistency can lead to variations in clarity and readability.
6. Based on the proposed Auxiliary Loss Construction method, there appears to be some degree of redundancy among the four losses defined in Eq. (8), particularly between L^T, L^S, and L^R.
7. In addition to retrieval performance, inference efficiency and model parameters are critical factors in the field of retrieval. The ability to deploy models effectively in real-world applications often hinges on these aspects. For example, Pic2word, which contains only one MLP as trainable parameters, highlights how simplicity can lead to enhanced efficiency without compromising performance.

1、In line 269, the structure of the Transformer encoder could be explained in the supplementary materials, or a reference could be added. 2、In the Methods section, there is no explanation of how $\mathcal{L}^{T}$ and $\mathcal{L}^{S}$ are derived; please provide this information in the Methods section. 3、In line 432, the title above Table 4 is incorrect; it is labeled as Figure 4. (Table 5; Table 7; Table 9). Please carefully review the manuscript for existing minor errors. Note: Please revise the Method section carefully, as it is crucial to the paper. Some implementation details that cannot be included in the main text can be added to the supplementary materials, which will help readers understand the underlying principles. I hold a positive attitude toward your work, but the manuscript requires thorough revision.

1. In the abstract section, the challenges of this composed video retrieval task are described too simplistically. The difficulty of fine-grained video-text modeling is a long-standing and intractable problem. The authors could provide a more nuanced description of the challenges.

2. The author should conduct a more detailed exploration of the temporal encoder. Dicosa[1] found that using the text-side [cls] token to perform cosine similarity with each video frame feature, and then passing through softmax to convert into a probability distribution to guide the weighted summation of video features is a better compression method. Could this process bring performance improvements in composed video retrieval?

3. The author lacks citations of relevant papers. Specifically, DPC-KNN has also been used in other video retrieval works [2], and relevant studies include Dicosa [2] and FreestyleRet [3]. FreestyleRet contains the exploration of composed image retrieval.

[1]. Text-video retrieval with Disentangled Conceptualization and Set-to-Set Alignment [2]. Video-Text as Game Players: Hierarchical Banzhaf Interaction for Cross-Modal Representation Learning [3]. FreestyleRet: Retrieving Images from Style-Diversified Queries

1. Unclear dataset construction details: There remain lots of missing details in the dataset construction. For example, how to handle the video when prompting, such as how many frames are used.

2. Quality control of the proposed dataset. In L160, the authors select the top 20 similar videos for each video based on the cosine similarity between them. It is unclear about the cosine similarity selection accuracy. The authors should clarify this through human mannual check or some other solutions.

3. The proposed method is similari to existing method [1]. These two papers share the similar fine-grained cross-modal alignment and fusion methodology. The authros should cite this paper and discuss the differences.

4. In experiments, the authors only include the results in the proposed dataset. It is unclear about the performance of the proposed method in general video-text retrieval datasets. It is recommended to include the discussion about it.

5. The overall paper writing can be polished. For example, the presented figures should be presented as vector images

[1] Disentangled Representation Learning for Text-Video Retrieval