Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
TGA: True-to-Geometry Avatar Dynamic Reconstruction
Authors: Bo Guo, Sijia Wen, Ziwei Wang, Yifan Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that TGA achieves superior geometric accuracy. Project page: https://superkeaibb.github.io/TGA/. ... We empirically demonstrate the advanced performance of the proposed method, demonstrating significant improvements in reconstruction accuracy, dynamic capability, training efficiency, and inference time. ... We evaluate our method on the Ne RSemble [33], Multiface [66] and NHA Dataset [42]. ... We compare TGA against our baselines on the Ne RSemble dataset [33] by reconstructing each avatar from 16-view RGB sequences and present qualitative results in Fig. 5. ... Additionally, we conduct a qualitative and quantitative evaluation of the reconstructed meshes on the Multiface Dataset [66] in Fig. 6. We focus on three metrics: L1-Chamfer distance, normal MAE (Mean Angular Error), and Recall [65] ... We present the comparison with Sur Fhead [4] and GHA [51] in Fig. 7 and Tab. 1, respectively. ... We also evaluate mesh extraction speed across different training iterations by reporting the number of mesh vertices at key checkpoints and the time required for mesh extraction per frame. |
| Researcher Affiliation | Academia | Bo Guo1 Sijia Wen1 Ziwei Wang1 Yifan Zhao2 1Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University 2State Key Laboratory of Virtual Reality Technology and Systems, SCSE&QRI, Beihang University EMAIL |
| Pseudocode | Yes | B Algorithm B.1 Pivot and Filter Hopping Points. Here, we represent the proposed algorithm for Pivot and Filter Hopping Points in Sec. 3.3.2. Algorithm 1: Pivot and Filter Hopping Points Input: T (BVH Tree), G (Gaussians), τ (Hop Threshold) Output: Updated T , hopping points H 1 function PIVOT(n, G, H, τ) 2 Refit AABB(G(n)); 3 if n is leaf then 4 Split Candidate(n); 5 else if n has two leaf children then 6 Merge(n); 7 else if n.depth 3 then 8 foreach ρ Lift & Reorder Rebalancers do 9 Compute SAH(n, ρ); 10 if Best Rotation(n) = NONE then 11 Apply Best Rotation(n), Refit AABB(G(n)), Compute Cost(n, ρ); 12 if Cost(n) > τ then 13 H H G(n); 14 function LRD(n, G, H, τ) 15 if n = NULL then 16 LRD(n.left, G, H, τ), LRD(n.right, G, H, τ); 17 PIVOT(n, G, H, τ); 18 foreach frame do 19 H , LRD(T , G, H, τ), Incremental Delaunay(H); |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will release our code after our paper gets accepted. |
| Open Datasets | Yes | Datasets. We evaluate our method on the Ne RSemble [33], Multiface [66] and NHA Dataset [42]. The Ne RSemble captures detailed facial dynamics, and the data is calibrated with sub-millimeter accurate camera poses and high-quality foreground segmentation. The Multiface dataset captures subjects covering dense multiview camera captures, rich facial expressions, and ground truth mesh to evaluate the 3D reconstruction efficiency. The NHA real dataset contains sequences that are suitable for the evaluation of full dynamic head approaches. We use it to evaluate the novel-view synthesis and self-reenactment rendering performance of our method. |
| Dataset Splits | No | The paper does not explicitly provide specific dataset split information (exact percentages, sample counts, or explicit instructions for creating custom splits). It mentions using |
| Hardware Specification | Yes | All experiments are performed on NVIDIA RTX 4090 GPUs, using the same hyperparameters as Gaussian Avatars [2]. |
| Software Dependencies | No | The paper mentions that the method is built upon 3DGS [1] and other related works, implying the use of associated software frameworks and libraries (e.g., PyTorch, CUDA), but it does not specify any version numbers for these software dependencies or other key libraries. |
| Experiment Setup | Yes | TGA converges within 300k iterations. All experiments are performed on NVIDIA RTX 4090 GPUs, using the same hyperparameters as Gaussian Avatars [2]. ... We set the hyperparameters following these works and λnr as 0.01. |