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..
Availability-aware Sensor Fusion via Unified Canonical Space
Authors: Dong-Hee Paek, SEUNG-HYUN KONG
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the K-Radar dataset demonstrate that ASF achieves improvements of 9.7% in APBEV (87.2%) and 20.1% in AP3D (73.6%) in object detection at Io U=0.5, while requiring a low computational cost. All codes are available at https://github.com/kaist-avelab/k-radar. |
| Researcher Affiliation | Academia | Dong-Hee Paek Seung-Hyun Kong CCS Graduate School of Mobility KAIST EMAIL |
| Pseudocode | No | The paper describes methods using mathematical formulations (Eqs. 1-9) and prose, but does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | All codes are available at https://github.com/kaist-avelab/k-radar. |
| Open Datasets | Yes | The effectiveness of our proposed ASF method has been validated on the K-Radar dataset (Paek et al., 2022), demonstrating improvements of 9.7% in APBEV (87.2%) and 20.1% in AP3D (73.6%) for detection performance at Io U=0.5 compared to state-of-the-art (SOTA) methods (Chae et al., 2024; Huang et al., 2025), which includes the performance in extreme situations such as sensor degradation or failure (i.e., unavailable). |
| Dataset Splits | No | For comparison with SOTA methods, we utilize two K-Radar benchmark variants. Benchmark v1.0 (Paek et al., 2022; Chae et al., 2024; Huang et al., 2025) focuses on the Sedan class within a driving corridor region of [0m, 72m] [-6.4m, 6m] (X Y Z). For ablation studies and qualitative analysis, we use benchmark v2.0, which covers a wider area [0m, 72m] [-16m, 16m] [-2m, 7.6m] and includes both Sedan and Bus or Truck classes. |
| Hardware Specification | Yes | We implement the ASF on a single RTX3090 GPU with 24GB VRAM. |
| Software Dependencies | No | We utilize established methods for the sensor-specific encoders and detection head. Specifically, we adopt BEVDepth (Li et al., 2023), SECOND (Yan et al., 2018), and RTNH (Paek et al., 2022) backbones for camera, Li DAR, and 4D Radar, respectively, along with a SSD detection head (Liu et al., 2016). ASF is trained for 11 epochs using Adam W (Loshchilov and Hutter, 2017) optimizer with a learning rate 0.001 and a batch size 2. |
| Experiment Setup | Yes | ASF is trained for 11 epochs using Adam W (Loshchilov and Hutter, 2017) optimizer with a learning rate 0.001 and a batch size 2. The voxel size for the fused FM is set to 0.4m, consistent with (Paek et al., 2022). |