Meta-Continual Learning of Neural Fields
Authors: Seungyoon Woo, Junhyeog Yun, Gunhee Kim
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive evaluations across image, audio, video reconstruction, and view synthesis tasks on six diverse datasets, demonstrating our method s superiority in reconstruction quality and speed over existing MCL and CL-NF approaches. |
| Researcher Affiliation | Academia | Seungyoon Woo, Junhyeog Yun, Gunhee Kim Seoul National University EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Modularized MCL-NF with Fisher Information Maximization Loss |
| Open Source Code | Yes | Code is available at https://github.com/seungyoon-woo/MCL-NF. |
| Open Datasets | Yes | We carry out extensive empirical evaluation across the image, audio, video reconstruction, and view synthesis tasks on six diverse datasets, including three 2D image datasets (Celeb A (Liu et al., 2015), FFHQ (Karras et al., 2019), and Image Nette (Deng et al., 2009)), one video dataset (Vox Celeb2 (Chung et al., 2018)), one audio dataset (Libri Speech (Panayotov et al., 2015)), and one Ne RF dataset (Matrix City (Li et al., 2023)). |
| Dataset Splits | Yes | We set the number of continual tasks to four, following prior works (Mi & Xu, 2023; Cho et al., 2022). This translates to four frames in the video domain and four 3D grids in the Ne RF domain. ...We partition the city blocks into meta-training and meta-test sets, allocating 70% of the total blocks to meta-training and the remaining 30% to meta-testing. |
| Hardware Specification | Yes | In this work, we evaluate the computational cost using our PyTorch implementation, conducted on NVIDIA TITAN X Pascal GPUs which have 12 GB of VRAM. |
| Software Dependencies | No | The paper mentions 'our PyTorch implementation' but does not specify a version number for PyTorch or any other software libraries. |
| Experiment Setup | Yes | We set the number of continual tasks to four... We run all experiments three times, reporting average performance until 500K outer steps... The coordinate-based MLP consists of five layers with Sine activations, configured with d = 128, din = 2 for image signal input, and dout = 3 for RGB outputs... Following standard Ne RF practices, we employ a ten-layer MLP with ReLU activations, configured with d = 1024, din = 5 for (x, y, z, θ, ϕ) coordinates, where θ and ϕ represent viewing angles, and dout = 4 for RGB and density outputs. |