Progressive Fourier Neural Representation for Sequential Video Compilation
Authors: Haeyong Kang, Jaehong Yoon, DaHyun Kim, Sung Ju Hwang, Chang D. Yoo
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our PFNR method on the UVG8/17 and DAVIS50 video sequence benchmarks and achieve impressive performance gains over strong continual learning baselines. The PFNR code is available at https://github.com/ihaeyong/PFNR.git. Our method is validated on benchmark datasets for Video Task-incremental Learning (VTL) and compared against various continual learning baselines. In all experiments conducted for this paper, we utilize a multi-head configuration for continual video representation learning. |
| Researcher Affiliation | Academia | Haeyong Kang, Jaehong Yoon, Da Hyun Kim, Sung Ju Hwang, and Chang D. Yoo Korea Advanced Institute of Science and Technology (KAIST) {haeyong.kang, jaehong.yoon, dahyun.kim, sjhwang82, cd_yoo}@kaist.ac.kr |
| Pseudocode | Yes | Algorithm 1 Progressive Fourier Neural Representation (PFNR) for VCL |
| Open Source Code | Yes | The PFNR code is available at https://github.com/ihaeyong/PFNR.git. |
| Open Datasets | Yes | We validate our PFNR method on the UVG8/17 and DAVIS50 video sequence benchmarks... The UVG dataset videos are available online at https://ultravideo.fi/#main. ... the Densely Annotation Video Segmentation dataset (DAVIS) (Perazzi et al., 2016). |
| Dataset Splits | No | The paper mentions 'validation' in the context of general model evaluation and loss functions but does not provide specific details about dataset splits (e.g., percentages, sample counts) for training, validation, or test sets. For example, it says 'We validate our PFNR method on the UVG8/17 and DAVIS50 video sequence benchmarks' and 'We evaluate the video quality, average video session quality, and backward transfer with two metrics...' but does not specify how data was partitioned for validation. |
| Hardware Specification | Yes | All experiments are run with NVIDIA RTX8000. ... We train and test two baselines (Ne RV, ESMER) with f-Ne RV2 using one GPU (TITAN V, 12G)... |
| Software Dependencies | No | The paper states, 'We implement our model in Py Torch and train it in full precision (FP32).' However, it does not provide specific version numbers for PyTorch or any other software libraries or dependencies used, which are necessary for reproducible descriptions. |
| Experiment Setup | Yes | We train WSN, PFNR, Ne RV (STL), and MTL using Adam optimizer with a learning rate 5e-4. For the ablation study on UVG8 and UVG17, we use a cosine annealing learning rate schedule (Loshchilov & Hutter, 2016), batch size of 1, training epochs of 150, and warmup epochs of 30 unless otherwise denoted. |