ShapeBoost: Boosting Human Shape Estimation with Part-Based Parameterization and Clothing-Preserving Augmentation

Authors: Siyuan Bian, Jiefeng Li, Jiasheng Tang, Cewu Lu

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our method outperforms other state-of-the-art methods in diverse body shape situations as well as in varied clothing situations.
Researcher Affiliation Collaboration Siyuan Bian1, Jiefeng Li1, Jiasheng Tang3,4, Cewu Lu1,2 1Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China 2Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China 3DAMO Academy, Alibaba group, Hangzhou, China 4Hupan Lab, Hangzhou, China {biansiyuan, ljf likit, lucewu}@sjtu.edu.cn, jiasheng.tjs@alibaba-inc.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an unambiguous statement of code release or a link to a code repository for the described methodology.
Open Datasets Yes We use 3DPW (von Marcard et al. 2018), Human3.6M (Ionescu et al. 2013), COCO (Lin et al. 2014), AGORA (Patel et al. 2021) and Model Agency Dataset (Choutas et al. 2022) for training.
Dataset Splits Yes On HBW validation set, our method outperforms previous SOTA methods. We also make ablation studies with different training data quantitatively. The results are shown in Tab. 6. When the data augmentation module is not used, the performance of our model drops on both HBW and SSP-3D dataset. This shows the effectiveness of our data augmentation module.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4).
Experiment Setup Yes The overall loss of our pipeline is formulated as L = Lpose + µ2Ldecomp + µ3Lshape. In shape loss, we supervise the predicted part widths predicted by the CNN backbone. Specifically, we require the projection results of the part slice widths and the vertex widths to be close to the target value after data augmentation. K represents the number of vertices in the human mesh model and J represents the number of joints. j ˆw2D j w2D j 2 2 + µ0 k ˆw2D k w2D k 2 2. Shape-decompose loss ensures that the shape reconstruction module predicts a valid human mesh while best preserving the part slice widths and bone lengths predicted by the CNN backbone. It consists of three loss functions Ldecomp = Lbone + Lwidth + µ1Lreg