Deductive Learning for Weakly-Supervised 3D Human Pose Estimation via Uncalibrated Cameras
Authors: Xipeng Chen, Pengxu Wei, Liang Lin1089-1096
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On three 3D human pose benchmarks, we conduct extensive experiments to evaluate our proposed method, which achieves superior performance in comparison with state-of-the-art weak-supervised methods. Particularly, our model shows an appealing potential for learning from 2D data captured in dynamic outdoor scenes, which demonstrates promising robustness and generalization in realistic scenarios. |
| Researcher Affiliation | Collaboration | Xipeng Chen1, Pengxu Wei1*, Liang Lin1, 2 1 Sun Yat-Sen University 2 Dark Matter AI Research |
| Pseudocode | No | No explicitly labeled 'Pseudocode' or 'Algorithm' block was found. The paper describes the method using mathematical formulas and text. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/Xipeng Chen/DWSL-3D-pose. |
| Open Datasets | Yes | Our experiments have been conducted on three 3D human pose datasets, i.e., Human3.6M (Ionescu et al. 2013), MPI-INF-3DHP dataset (Mehta et al. 2017a), and Ski-Pose PTZ (Rhodin et al. 2018). |
| Dataset Splits | Yes | Human3.6M (Ionescu et al. 2013) consists of 3.6 million images and has 11 subjects with 15 daily actions from 4 different viewpoints. Following the standard protocol in (Martinez et al. 2017), subject 1, 5, 6, 7, 8 are for training and subject 9, 11 are for evaluation. |
| Hardware Specification | Yes | We train the network on an RTX 2080 GPU with a batchsize of 64 for 100 epochs. |
| Software Dependencies | No | The paper mentions using a 'trained 2D pose estimator (Duan et al. 2019)' and 'Open Pose (Cao et al. 2018)' for 2D detections, and that 'Our code is publicly available at https://github.com/Xipeng Chen/DWSL-3D-pose.' However, it does not provide specific version numbers for any libraries or frameworks like Python, PyTorch, TensorFlow, etc., that would be needed for replication. |
| Experiment Setup | Yes | 2D poses used in our model are normalized to unit size together with their pelvis points at origin. For the input 2D poses, they are augmented by a random rotation within 30 , and are re-scaled by a factor within 1 0.1. For the target 2D poses, we introduce the flip operation to generate data as if it was taken from a virtual camera, via multiplying the x coordinate by -1. In every training epoch, we randomly choose a pair of viewpoints for every frame when there exist multiply ones. The weights in the losses, (αrec, αpos, αsym, αang), are set to (1, 1, 10, 10 3). We train the network on an RTX 2080 GPU with a batchsize of 64 for 100 epochs. |