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..

CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction

Authors: Yiyi Liu, Chunyang Liu, Bohan Wang, Weiqin Jiao, Bojian Wu, Lubin Fan, Yuwei Chen, Fashuai Li, Biao Xiong

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Structured3D and Scene CAD show that CAGE achieves state-of-the-art performance, with F1 scores of 99.1% (rooms), 91.7% (corners), and 89.3% (angles). The method also demonstrates strong cross-dataset generalization, underscoring the efficacy of our architectural innovations.
Researcher Affiliation Academia 1Wuhan University of Technology 2University of Twente 3Independent Researcher 4Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences 5The Advanced Laser Technology Laboratory of Anhui Province *Corresponding Authors: EMAIL, EMAIL
Pseudocode No The paper describes the CAGE framework and its components, including the dual-query transformer decoder and loss formulation, but it does not include a clearly labeled pseudocode or algorithm block with structured steps.
Open Source Code No Our code and pre-trained models will be released upon acceptance.
Open Datasets Yes We evaluate our method on two large-scale indoor datasets, including Structured3D [16] and Scene CAD [17].
Dataset Splits Yes Following previous work [13; 12], we split Structured3D into 3,000 training, 250 validation, and 250 test samples. For Scene CAD, we use the provided splits of 828 training and 127 validation samples.
Hardware Specification Yes Our model is implemented in Py Torch and trained on a single NVIDIA RTX 4090 GPU.
Software Dependencies No Our model is implemented in Py Torch and trained on a single NVIDIA RTX 4090 GPU. We use Swin Transformer V2 [45] as the default image backbone, while Res Net50 [46] and Swin Transformer V1 [47] are used in ablation studies. ... We optimize using Adam [48].
Experiment Setup Yes We train our model using the Adam optimizer [48] with a weight decay of 1e 4. Depending on the dataset size, we train the model on Structured3D for 650 epochs with an initial learning rate of 2e 4, and on Scene CAD for 400 epochs with an initial learning rate of 5e 5. In both cases, the learning rate is decayed by a factor of 0.1 during the final 20% of epochs. The loss weights are set as λcls = 0.6, λedge = 6, λras = 1, λDN_cls = 0.6, and λDN_edge = 6. For perturbed queries, the noise scale is set to 0.2 for class and 0.4 for edge. We set the number of room feature codes m to 20 and the number of edge queries n to 40.