PoseGen: Learning to Generate 3D Human Pose Dataset with NeRF

Authors: Mohsen Gholami, Rabab Ward, Z. Jane Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments show that the proposed Pose Gen improves two baseline models (SPIN and Hybr IK) on four datasets with an average 6% relative improvement.
Researcher Affiliation Academia Mohsen Gholami, Rabab Ward, Z. Jane Wang University of British Columbia, Vancouver, Canada {mgholami, rababw, zjanew}@ece.ubc.ca
Pseudocode No The paper describes the framework and components using textual descriptions and mathematical formulations, but no structured pseudocode or algorithm blocks are provided.
Open Source Code Yes Code is available at https: //github.com/mgholamikn/Pose Gen.
Open Datasets Yes We use three datasets that are not used during pretraining of pose estimator P, including 3DPW, AGORA, and SKI-Pose, for evaluation. To further evaluate the effectiveness of the proposed Pose Gen on boosting the performance of P on IND datasets used in the pre-training procedure, we also perform an evaluation on 3DHP. ... 3DPW (von Marcard et al. 2018) includes in-the-wild images... AGORA (Patel et al. 2021) uses 4240 high-quality textured scans of people... SKI-Pose (Sp orri 2016) is captured in a ski resort... MPI-INF-3DHP (3DHP) includes data from 8 subjects.
Dataset Splits No The paper mentions using training and test sets but does not provide specific details on validation splits (e.g., percentages or sample counts for validation data) for any of the datasets.
Hardware Specification No The paper does not provide specific details regarding the hardware used for running the experiments (e.g., GPU models, CPU models, memory specifications).
Software Dependencies No The paper mentions using pre-trained models like Hybr IK and SPIN, and a Ne RF model (A-Ne RF), but does not specify software versions for these tools or for the underlying deep learning frameworks (e.g., PyTorch, TensorFlow) or other dependencies like Python or CUDA versions.
Experiment Setup Yes The overall loss of G is LG = w1Ladv + w2Lfb, (8)... In the above formula, d is a threshold to exclude samples with large errors." and "In Scenario 1, the objective of G is to generate plausible poses while minimizing the training loss of the pose estimator P: min G,P max D L(G, D, P). (1) In Scenario 2 the objective of G is to generate plausible poses while increasing the training loss of the pose estimator P: min G max D,P L(G, D, P). (2)