Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Repurposing Marigold for Zero-Shot Metric Depth Estimation via Defocus Blur Cues
Authors: Chinmay Talegaonkar, Nikhil Gandudi Suresh, Zachary Novack, Yash Belhe, Priyanka Nagasamudra, Nicholas Antipa
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare our method against existing state-of-the-art zero-shot MMDE methods on a self-collected real dataset, showing quantitative and qualitative improvements. Our implementation is available at https://github.com/chinmay0301ucsd/Diffusion Cam. |
| Researcher Affiliation | Academia | Chinmay Talegaonkar , Nikhil Gandudi Suresh, Zachary Novack, Yash Belhe, Priyanka Nagasamudra, Nicholas Antipa University of California San Diego Corresponding author: EMAIL |
| Pseudocode | No | The paper describes the method using equations and text descriptions, but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Our implementation is available at https://github.com/chinmay0301ucsd/Diffusion Cam. |
| Open Datasets | Yes | We collect a dataset of 7 diverse real-world indoor scenes, captured at different defocus blur levels. ... Standard monocular depth datasets [49, 35] typically capture in-focus images at a single aperture per scene, making them unsuitable for evaluating our method. ... We evaluate our method against some of the recently popular MMDE methods Uni Depth [39], Metric3D [74], and MLPro [7] on our collected dataset (table 2) and the NYUv2 [35] test set. ... Depth Anythingv2 [71] fine-tuned on the Hypersim[42] dataset (row 1), NYU-v2 train set (row 2)... We will release the self-collected dataset in the supplementary material with guidelines on how to use it, and all the camera parameters used while capturing it. |
| Dataset Splits | Yes | We evaluate our method against some of the recently popular MMDE methods Uni Depth [39], Metric3D [74], and MLPro [7] on our collected dataset (table 2) and the NYUv2 [35] test set. ... Depth Anythingv2 [71] fine-tuned on the Hypersim[42] dataset (row 1), NYU-v2 train set (row 2)... |
| Hardware Specification | Yes | We run the optimization for 200 iterations, which takes roughly 3.5-4 minutes on an NVIDIA A-40 GPU with peak memory usage of 15 GB. |
| Software Dependencies | No | The paper mentions 'PyTorch' and 'CUDA kernels' but does not specify their version numbers or other software dependencies with specific versions. |
| Experiment Setup | Yes | We run the optimization for 200 iterations... We use the Adam optimizer with a learning rate of 1.5 10 3 for z(d) T , 5 10 3 for (a, b), and default values for optimizer parameters. Note that we use the same scene bounds smin = 1.49, smax = 3.5 for all the real scenes in our dataset. |