bit2bit: 1-bit quanta video reconstruction via self-supervised photon prediction
Authors: Yehe Liu, Alexander Krull, Hector Basevi, Ales Leonardis, Michael Jenkins
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method using both simulated and real data. On simulated data from a conventional video, we achieve 34.35 mean PSNR with extremely photon-sparse binary input (<0.06 photons per pixel per frame). We also present a novel dataset containing a wide range of real SPAD high-speed videos under various challenging imaging conditions. and 4 Experiments and 5 Results and discussions |
| Researcher Affiliation | Collaboration | Yehe Liu1,2 Alexander Krull3, Hector Basevi3 Aleš Leonardis3 Michael Jenkins1,2, 1Case Western Reserve University 2Opsi Clear LLC 3University of Birmingham |
| Pseudocode | No | The paper describes the method verbally and provides figures for network architecture, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source code of the parameterized model with bit2bit: https://github.com/lyehe/ssunet |
| Open Datasets | Yes | Simulated and real SPAD data: https://drive.google.com/drive/folders/1M5bsmsa LBk Ym O7n MUj K5_m71Ron Op-P9 and Additionally, we use a real video with 100k frames published in [8]. |
| Dataset Splits | No | The paper mentions 'increasing validation loss' (Table 1, S16) indicating that validation was performed, but it does not specify the train/validation/test dataset splits (e.g., percentages or sample counts) for the overall dataset used in the experiments. |
| Hardware Specification | Yes | We primarily used Nvidia 3090/4090 for training... and The inference speed is above 3 volumes per second (150 fps) on a NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions software components like '3D Res UNet', 'ADAMW optimizer', 'Group normalization', 'Ge LU', and 'pixel shuffling', but does not provide specific version numbers for any of these libraries or the underlying programming language/framework. |
| Experiment Setup | Yes | Models were trained using the ADAMW optimizer for 150 epochs, with 250 steps per epoch and 4 batches of random crops of 32x256x256 (TXY) per step [36]. and Table S1: The baseline hyperparameters used in training. This produces the PSNR 33.93/0.99 and SSIM 0.959/0.007 with the simulated data. ... Start features 32 Depth 5 ... Group norm 8 ... Batch size 4 Epochs 150 Optimizer adamw Learn Rate 0.00032 |