UP-NeRF: Unconstrained Pose Prior-Free Neural Radiance Field

Authors: Injae Kim, Minhyuk Choi, Hyunwoo J. Kim

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments verify the superior performance of our method compared to the baselines including BARF and its variants in a challenging internet photo collection, Phototourism dataset.
Researcher Affiliation Academia Injae Kim Korea University dna9041@korea.ac.kr Minhyuk Choi Korea University sodlqnf123@korea.ac.kr Hyunwoo J. Kim Korea University hyunwoojkim@korea.ac.kr
Pseudocode No The paper describes its methods using mathematical equations and textual explanations, but it does not include a formally labeled 'Algorithm' or 'Pseudocode' block.
Open Source Code Yes The code of UP-Ne RF is available at https://github.com/mlvlab/UP-Ne RF.
Open Datasets Yes We report results on the Phototourism dataset. It consists of internet photo collections of famous landmarks and we select 4 scenes, Brandenburg Gate, Sacre Coeur, Taj Mahal, and Trevi fountain, which are also used in Ne RF-W.
Dataset Splits Yes We follow the split used by Ne RF-W [18] and downsample each image by 2 times.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU model, CPU type) used for running the experiments.
Software Dependencies No The paper mentions using specific software components like 'Adam optimizer', 'Adam W', 'Vi T', 'DPT', and 'DINO model' but does not provide specific version numbers for any of them.
Experiment Setup Yes All the models are trained for 600K iterations with randomly sampled 2048 pixel rays at each step with a learning rate of 5 10 4 decaying to 5 10 5 for Ne RF and transient network T , and 2 10 3 decaying to 1 10 3 for pose p and two factors si and ti of depth. We use Adam optimizer [51] across all the experiments except test-time appearance optimization, where Adam W [52] is used instead. The number of sampling points in each ray for volumetric rendering is set to 128 for both coarse and fine models. We use the default coarse-to-fine strategy of BARF which starts from training progress 0.1 to 0.5. We set the scheduling parameters u and v to be 0.1 and 0.5, respectively, the same as the parameters of coarse-to-fine. The hyperparameter λd for depth loss is set to 0.001. We set βmin to 0.1 and λα to 1.0.