PaperHub
4.3
/10
withdrawn3 位审稿人
最低3最高5标准差0.9
5
3
5
4.3
置信度
正确性2.3
贡献度2.0
表达2.3
ICLR 2025

Textual $\textbf{D}$ecomposition then Sub-motion-space $\textbf{S}$cattering for $\textbf{O}$pen-Vocabulary Motion Generation

OpenReviewPDF
提交: 2024-09-13更新: 2024-11-13

摘要

Text-to-motion generation is a crucial task in computer vision, which generates the target 3D motion by the given text. The existing annotated datasets are limited in scale, resulting in most existing methods overfitting to the small datasets and unable to generalize to the motions of the open domain. Some methods attempt to solve the open-vocabulary motion generation problem by aligning to the CLIP space or using the Pretrain-then-Finetuning paradigm. However, the current annotated dataset's limited scale only allows them to achieve mapping from sub-text-space to sub-motion-space, instead of mapping between full-text-space and full-motion-space (full mapping), which is the key to attaining open-vocabulary motion generation. To this end, this paper proposes to leverage the atomic motion (simple body part motions over a short time period) as an intermediate representation, and leverage two orderly coupled steps, i.e., Textual Decomposition and Sub-motion-space Scattering, to address the full mapping problem. For Textual Decomposition, we design a fine-grained description conversion algorithm, and combine it with the generalization ability of a large language model to convert any given motion text into atomic texts. Sub-motion-space Scattering learns the compositional process from atomic motions to the target motions, to make the learned sub-motion-space scattered to form the full-motion-space. For a given motion of the open domain, it transforms the extrapolation into interpolation and thereby significantly improves generalization. Our network, $DSO$-Net, combines textual $d$ecomposition and sub-motion-space $s$cattering to solve the $o$pen-vocabulary motion generation. Extensive experiments demonstrate that our DSO-Net achieves significant improvements over the state-of-the-art methods on open-vocabulary motion generation.
关键词
Motion GenerationOpen-Vocabulary

评审与讨论

审稿意见
5

In this paper, the authors proposed DSO-Net, an open-vocabulary text-to-motion generation framework. The proposed method achieve competitive performance across multiple benchmark.

优点

  1. The proposed method achieve competitive performance across multiple benchmark.

缺点

  1. The proposed atomic motion idea has been discover in [1].
  2. As claimed in the main paper that human beings tend to partition it into the combination of several simple body part motions over a short time period, there are no quantitative or qualitative results to support the claim.
  3. Missing metrics. MMDist and MModality are commonly used metric in text2motion task which are not included in this paper.

[1] Language-guided human motion synthesis with atomic actions, Zhai et al, ACM MM

问题

N/A

审稿意见
3

This paper proposes a text-to-motion framework, aimed at the open-vocabulary motion generation problem. Specifically, this paper proposes to learn a sub-motion space, using atomic motion representations as an intermediate representation, and leveraginhg textual decomposition and sub-motion space scattering for the action representation mapping learning. The proposed method achieves strong performance over existing baselines.

优点

  1. The motivation to decompose an action description into a set of atomic action is strong. Using LLMs to extract the atomic actions in the textual format has not been widely used before.
  2. The proposed method achieves strong performances, especially the R-Precision that reflects motion-text alignment, on common benchmarks.

缺点

  1. This paper is not well-written. One of the most fundamental problem is how the motion is generated and the training objectives are not described. Do the authors use diffusion models or VAE?
  2. While the the concept of text-based atomic action decomposition is well-motivated, it lacks one of the most important comparison results, i.e., the comparison with CLIP-based alignment (top right of Fig. 1). Without this, it is unclear whether the text-based decomposition outperforms the decomposition in the latent space.
  3. This paper lacks a direct comparison with the latent-space atomic action decomposition method [a], which also learns atomic actions using cross-attention. The authors should describe the differences and compare the performance.
  4. The proposed method could not achieve strong FID on common benchmarks, indicating the generated motions are not realistic, which could be verified in some of the videos in the supplement.

[a] Zhai, Yuanhao, Mingzhen Huang, Tianyu Luan, Lu Dong, Ifeoma Nwogu, Siwei Lyu, David Doermann, and Junsong Yuan. "Language-guided human motion synthesis with atomic actions." In Proceedings of the 31st ACM International Conference on Multimedia, pp. 5262-5271. 2023.

问题

  1. Can the authors verify why the text-based action decomposition outpeforms the decomposition in the latent-space?
  2. Can the authors explain why the proposed method achieves strong R-Precision, but could not improve over FID?
审稿意见
5

This paper proposes a method for improving open-vocabulary motion generation. It consists of two main components, i.e., textual decomposition and sub-motion-space scattering, where the former uses LLM to generate the descriptions of atomic motions and the latter further learns the generative combination of atomic motions by a text-motion alignment (TMA) module and a compositional feature fusion (CFF) module.

优点

  1. The idea of atomic actions has been explored by some existing works (see weakness). Specifically, the paper
    decomposes the motion text into the texts of different body parts.

  2. The paper is well-structured and provides a clear and detailed description of the proposed framework.

缺点

1 Limited novelty. The idea of atomic texts/actions has been proposed in [A] and [B], but the authors did not cite these papers. The differences between the proposed method and them should be fully discussed.

[A] Language-guided Human Motion Synthesis with Atomic Actions, ACM MM 2023

[B] Generative Action Description Prompts for Skeleton-based Action Recognition, ICCV 2023

2 Insufficient experiments. The paper fails to compare the proposed motion generation framework with some important SOTA methods, e.g. [C], [D]…

[C] AttT2M: Text-Driven Human Motion Generation with Multi-Perspective Attention Mechanism, ICCV 2023

[D] MMM: Generative Masked Motion Model, CVPR 2024

问题

1 It would be helpful to compare with more state-of-the-art motion generation algorithms.

2 Could you please provide a brief discussion on the computational complexity, which is important to real-time applications? And discuss any limitations or potential pitfalls in the proposed methods?

撤稿通知

withdrawal improves this work.