TSDF-Based Efficient Motion-Compensated Temporal Interpolation for 3D Dynamic Sequences
Authors: Soowoong Kim, Minseong Kwon, Junho Choi, Gun Bang, Seungjoon Yang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments and Results Experimental Setup Datasets The study employed a diverse set of dynamic mesh datasets for training and evaluation: the MPEG-curated dynamic mesh dataset (Xu, Lu, and Wen 2021), MITAMA (Vlasic et al. 2008), and 4DHuman Outfit (4DHO) (Armando et al. 2023). For evaluation, we selected three sequences (Longdress, Loot, Soldier) from the MPEG dataset, three sequences (Swing, Crane, Handstand) from the MITAMA dataset, and four sequences (deb-jea-walk, leo-tig-jump, ray-own-torso, ted-opt-side) from the 4DHO dataset. The remaining sequences from each dataset were used for training: seven sequences from the MPEG dataset, seven from MITAMA, and nine from 4DHO. Prior to training, the dynamic mesh data was preprocessed by converting the meshes into TSDF volumes. This involved partitioning the bounding box into a 128 128 128 grid and computing the signed distance from each voxel query point to the surface, truncated to the range of [ 1, 1]. Detailed information regarding these datasets is provided in the supplementary material. Baselines We compared our approach with several state-of-the-art models across various 3D representation techniques. Metrics We quantitatively evaluated our method from two perspectives. First, we measured geometric distance using Chamfer Distance (CD) (Fan, Su, and Guibas 2017) and Earth Mover s Distance (EMD) (Rubner, Tomasi, and Guibas 2000) when comparing with Neural PCI. ... Ablation Study We conducted ablations to verify the effectiveness of intermediate motion estimation, hybrid upsampling, and adaptive merging in enhancing performance. |
| Researcher Affiliation | Academia | Soowoong Kim1 , Minseong Kwon2 , Junho Choi2, Gun Bang1, Seungjoon Yang2 1Electronics and Telecommunications Research Institute 2Ulsan National Institute of Science and Technology EMAIL, EMAIL |
| Pseudocode | No | The paper describes the method using prose and network architecture diagrams (Figure 2 and Figure 3) but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Datasets The study employed a diverse set of dynamic mesh datasets for training and evaluation: the MPEG-curated dynamic mesh dataset (Xu, Lu, and Wen 2021), MITAMA (Vlasic et al. 2008), and 4DHuman Outfit (4DHO) (Armando et al. 2023). |
| Dataset Splits | Yes | For evaluation, we selected three sequences (Longdress, Loot, Soldier) from the MPEG dataset, three sequences (Swing, Crane, Handstand) from the MITAMA dataset, and four sequences (deb-jea-walk, leo-tig-jump, ray-own-torso, ted-opt-side) from the 4DHO dataset. The remaining sequences from each dataset were used for training: seven sequences from the MPEG dataset, seven from MITAMA, and nine from 4DHO. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. It only generally discusses computational power requirements for 3D scanning devices and performance metrics without detailing the experimental setup hardware. |
| Software Dependencies | No | The paper mentions several deep learning models and architectures (e.g., RAFT, FlowNet2.0, PointNet++) and algorithms (Marching Cubes), but it does not specify any software libraries, frameworks, or their version numbers used for implementation (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In our experiments, we set the hyperparameters as follows: N = 12, αf = 1 2N , αi = 1 2, αm = 1, and γ = 0.8. |