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..

Layer-Wise Modality Decomposition for Interpretable Multimodal Sensor Fusion

Authors: Jaehyun Park, Konyul Park, Daehun Kim, Junseo Park, Jun Won Choi

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate LMD on pretrained fusion models under camera radar, camera Li DAR, and camera radar Li DAR settings for autonomous driving. Its effectiveness is validated using structured perturbation-based metrics and modality-wise visual decompositions, demonstrating practical applicability to interpreting high-capacity multimodal architectures.
Researcher Affiliation Academia Jaehyun Park Konyul Park Daehun Kim Junseo Park Jun Won Choi Seoul National University EMAIL EMAIL
Pseudocode No The paper describes the Layer-Wise Modality Decomposition (LMD) method through textual explanations, mathematical formulations, and diagrams (Figure 1), but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/detxter-jvb/ Layer-Wise-Modality-Decomposition.
Open Datasets Yes The nu Scenes dataset [19] is a large-scale, multimodal dataset designed for autonomous driving tasks.
Dataset Splits Yes It encompasses 1000 scenes, segmented into 700 for training, 150 for validation, and 150 for testing, each lasting about 20 seconds.
Hardware Specification Yes Our method was evaluated for 6019 iterations with a batch size of 1 on 4 x NVIDIA RTX 3090 GPUs, which required approximately 3 hours in total.
Software Dependencies No The paper does not explicitly provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks used in the experiments.
Experiment Setup Yes table 5 presents radar-camera and Li DAR-camera configuration details used for the Simple BEV framework. Parameter Radar / Li DAR Camera Backbone Res Net101 Fusion Operation Concat & Conv Sweeps 5 Input Size (224, 400) BEV Coordinate (200, 8, 200)