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..
Can NeRFs "See" without Cameras?
Authors: Chaitanya Amballa, Yu-Lin Wei, Sattwik Basu, Zhijian Yang, Mehmet Ergezer, Romit Roy Choudhury
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate Echo Ne RF, we train on 2.4 GHz Wi Fi signals from NVIDIA s Sionna simulator [14], with floorplans from the Zillow s Indoor Dataset (ZIND) [9]. Results show consistent improvement over baselines in terms of the estimated floorplan s Io U and F1 score. Qualitative results show visually legible floorplans without any post-processing. |
| Researcher Affiliation | Collaboration | 1University of Illinois Urbana Champaign 2Amazon |
| Pseudocode | No | The paper describes the model using mathematical equations and textual explanations, but it does not include a clearly labeled pseudocode block or algorithm section. |
| Open Source Code | No | Code: We plan to release our code, data, and baselines soon. In the meantime, Section E provides sufficient details to allow readers to reproduce our results, especially since our network components are simple MLPs. |
| Open Datasets | Yes | Floorplans are drawn from the Zillow Indoor Dataset [9]. In each floorplan, we use the A algorithm [13] to generate a walking trajectory that traverses all rooms. We use the NVIDIA Sionna RT [15, 14] a ray tracer for radio propagation modeling to compute the ground truth signal power (also known as received signal strength index (RSSI)). |
| Dataset Splits | Yes | We partition our dataset into an 80-20 split, using 80% of the data for model training, including baselines. |
| Hardware Specification | Yes | We train our models on NVIDIA A100 GPUs. |
| Software Dependencies | No | For Echo Ne RF s network, we employ a simple 8-layer MLP with a hidden dimension of 256 units. For each voxel vj, the outputs from the final layer are passed through a sigmoid activation to obtain the opacity Ī“, and through a Gumbel softmax [16] layer to sample the output normal Ļ from one of the possible KĻ orientations. ... We use the ADAM optimizer [17] with 1.0 4 learning rate. The paper mentions software components like MLP, Gumbel softmax, and ADAM optimizer but does not specify their version numbers or the versions of underlying libraries like PyTorch or TensorFlow. |
| Experiment Setup | Yes | We choose Ī»1 = Ī»2 = 0.01 for Lo S training followed by Ī»1 = Ī»2 = 0.1 for training the Echo Ne RF model. ... We use the ADAM optimizer [17] with 1.0 4 learning rate. ... To help optimization and to encourage sparsity of the number of reflections, we use only the top-k contributions (k=10) while training the reflection model. |