MetaPix: Few-Shot Video Retargeting
Authors: Jessica Lee, Deva Ramanan, Rohit Girdhar
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experiment on in-the-wild internet videos and images and show our approach improves over widely-used baselines for the task. (Abstract) Section 4 is titled "EXPERIMENTS" and discusses "DATASETS AND EVALUATION", "EVALUATING METAPIX", and "ABLATIONS". It includes quantitative metrics (SSIM, PSNR, MSE) and comparisons in Table 1 and Figures 4, 5, 6, 7, 8, 9. |
| Researcher Affiliation | Academia | Jessica Lee Deva Ramanan Rohit Girdhar Carnegie Mellon University |
| Pseudocode | Yes | Algorithm 1 Meta-learning for video re-targeting for the Pose2Im setup. (Section 3) |
| Open Source Code | No | We will release the Meta Pix source code for details. (Section 3, Implementation Details) The only provided link, https://github.com/NVIDIA/pix2pix HD/, is for a third-party implementation (Pix2Pix HD), not the authors' Meta Pix code. |
| Open Datasets | Yes | Additionally, we collect a set of 10 more dance videos from You Tube (distinct from the above 8), as our pre-training and meta-learning corpus. We provide the list of You Tube video IDs for both in the supplementary. (Section 4.1) |
| Dataset Splits | No | Similar to (Zhou et al., 2019), we split each of the 8 test videos into a training and test sequence in 0.85:0.15 ratio... (Section 4.1) The paper describes a training and testing split but does not explicitly mention a separate validation split or dataset for hyperparameter tuning. |
| Hardware Specification | Yes | We implement Meta Pix for the Pose2Im base model by building upon a public Pix2Pix HD implementation1 in Py Torch, and perform all experiments on a 4 TITAN-X or GTX 1080Ti GPU node. (Section 3, Implementation Details) |
| Software Dependencies | No | We implement Meta Pix for the Pose2Im base model by building upon a public Pix2Pix HD implementation1 in Py Torch... (Section 3, Implementation Details) While 'Py Torch' is mentioned, no specific version number for PyTorch or any other software library or tool is provided. |
| Experiment Setup | Yes | The generator consists of 16 convolutional and deconvolutional layers... During this pretraining stage, the model is trained on all of the training frames for 10 epochs using learning rate of 0.0002 and batch size of 8... When finetuning for personalization... we train the first T/2 iterations using a constant learning rate of 0.0002, and the remaining iterations using a linear decay to 0... The batch size is fixed to 8... For the metalearning, we set the meta learning rate, E = 1 with a linear decay to 0, and train 300 meta-iterations. (Section 3, Implementation Details) |