LiftedCL: Lifting Contrastive Learning for Human-Centric Perception
Authors: Ziwei Chen, Qiang Li, Xiaofeng Wang, Wankou Yang
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate that Lifted CL outperforms state-of-the-art self-supervised methods on four human-centric downstream tasks, including 2D and 3D human pose estimation (0.4% m AP and 1.8 mm MPJPE improvement on COCO 2D pose estimation and Human3.6M 3D pose estimation), human shape recovery and human parsing. |
| Researcher Affiliation | Collaboration | Ziwei Chen1,2, Qiang Li3 , Xiaofeng Wang4, Wankou Yang1,2 1 School of Automation, Southeast University 2 Key Lab of Measurement and Control of Complex Systems of Engineering, Ministry of Education, Southeast University, Nanjing, China 3 Y-tech, Kuaishou Technology 4 Institute of Automation, Chinese Academy of Sciences |
| Pseudocode | No | The paper does not contain any sections or figures explicitly labeled as "Pseudocode" or "Algorithm". |
| Open Source Code | No | The paper states: "The training code of our framework is modified from the official Py Torch implementation of Mo Co1. 1https://github.com/facebookresearch/moco", but does not explicitly state that their own developed code is made open-source or provide a link to it. |
| Open Datasets | Yes | Our pre-training experiments are conducted on MS COCO (Lin et al., 2014) and only the training set is used for pre-training. |
| Dataset Splits | Yes | The fine-tuning is conducted on COCO train2017 dataset for 140 epochs with a mini-batch size of 128, including 57K images and 150K person instances. The evaluation is conducted on the val2017 set, containing 5000 images. |
| Hardware Specification | No | The paper mentions training "on 4 GPUs" but does not specify the model or type of GPUs, CPU, or any other specific hardware components used. |
| Software Dependencies | No | The paper mentions using "Py Torch" but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | We adopt SGD as the optimizer with initial learning rate of 0.03 and we set its weight decay and momentum to 0.0001 and 0.9. Each pre-training model is optimized on 4 GPUs with a cosine learning rate decay schedule and a mini-batch size of 128 for 200 epochs (details in appendix C). For both the invariant and equivariant contrastive learning, the dictionary size is set to 16384. The momentum is set to 0.999. Shuffling BN (He et al., 2020) is used during training. The temperature τ in contrastive loss is set to 0.2. |