Behavioral Recognition of Skeletal Data Based on Targeted Dual Fusion Strategy
Authors: Xiao Yun, Chenglong Xu, Kevin Riou, Kaiwen Dong, Yanjing Sun, Song Li, Kevin Subrin, Patrick Le Callet
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conducted experimental evaluations of the proposed FRF-GCN model on three large-scale datasets. We also performed extensive ablation experiments to validate the effectiveness of the proposed components. |
| Researcher Affiliation | Academia | Xiao Yun1, Chenglong Xu1, Kevin Riou2, Kaiwen Dong1*, Yanjing Sun1, Song Li1, Kevin Subrin2, Patrick Le Callet2 1School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, 221116, China 2Nantes Universit e, Ecole Centrale Nantes, CNRS, LS2N, UMR 6004, Nantes, France {xyun, clongxu, dongkaiwen, yjsun, lisong}@cumt.edu.cn, {kevin.riou, kevin.subrin, patrick.lecallet}@univ-nantes.fr |
| Pseudocode | No | The paper includes architectural diagrams and mathematical equations but no structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at: https://github.com/sunbeam-kkt/FRF-GCN-master. |
| Open Datasets | Yes | NTU RGB+D 60 dataset (Shahroudy et al. 2016), which is a large-scale skeleton dataset used for human action recognition models. It contains a total of 56,880 action sequences, 60 action categories, performed by different people, of which 40 actions are daily behaviors, 9 health-related actions, and 11 two-person interaction behaviors. The dataset is evaluated using two different evaluation protocols: Cross-Subject and Cross-View. NTU RGB+D 120 The NTU RGB+D 120 dataset (Liu et al. 2019) is a supplement to the NTU RGB+D 60 dataset. Kinetics-Skeleton 400 Kinetics-Skeleton 400 (Carreira and Zisserman 2017) is a very large dataset for behavior recognition. |
| Dataset Splits | Yes | The dataset is evaluated using two different evaluation protocols: Cross-Subject and Cross-View. In the Cross-Subject evaluation, the 106 participants are divided into training and testing sets, with each set containing 53 subjects. In the Cross-Setup evaluation, action sequences from even-numbered setup IDs are used for training, while action sequences from oddnumbered setup IDs are used for testing. |
| Hardware Specification | Yes | All experiments were conducted on RTX 3080 GPUs. |
| Software Dependencies | No | The paper mentions "Py Torch deep learning framework" but does not specify a version number or list other software dependencies with their versions. |
| Experiment Setup | Yes | The stochastic gradient descent (SGD) with Nesterov momentum (0.9) was employed as the optimization strategy, with a batch size of 56. The cross-entropy function was selected as the loss function for backpropagation gradients. A weight decay of 0.0003 and a learning rate of 0.1 were set. For NTU60,120 and Kinetics-Skeleton, the total epochs were set to 50, 60, and 65, respectively. The two-phase auto-attenuation renditions were 30, 40, 30, 50, 45, and 55, respectively. Unlike most of the previous studies, FRF-GCN uses a preprocessing method of data interpolation to supplement the skeleton sequence to 300 frames. Defaults to two individuals per sample, supplemented with zeros if there is only one. Additionally, a warm-up strategy (He et al. 2016) was applied during the first 5 epochs to enhance training stability. |