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..

RobIA: Robust Instance-aware Continual Test-time Adaptation for Deep Stereo

Authors: Jueun Ko, Hyewon Park, Hyesong Choi, Dongbo Min

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that Rob IA achieves superior adaptation performance across dynamic target domains while maintaining computational efficiency. 5 Experiments. Datasets. Evaluation Metrics. Main Results. Analyses. Ablation Studies.
Researcher Affiliation Academia 1Ewha Womans University 2Soongsil University EMAIL EMAIL EMAIL
Pseudocode No The paper includes Figure 1: "The Overview of Rob IA", which illustrates the model architecture and data flow, but it does not contain a dedicated pseudocode or algorithm block with structured, step-by-step instructions in a text format.
Open Source Code Yes https://github.com/0ju-un/RobIA
Open Datasets Yes To simulate TTA and CTTA scenarios, following prior work [8], all experiments were conducted on well-renowned stereo benchmarks, including KITTI RAW [40], Driving Stereo [41], and DSEC [42]. These datasets cover various conditions, such as different weather scenarios and urban cityscapes in both daylight and nighttime. The synthetic Flyingthings3D, part of the synthetic Scene Flow dataset [2] was used to pretrain the stereo model before test time.
Dataset Splits No The paper describes how the CTTA benchmark was constructed by sampling 500 frames per domain from existing TTA sequences and repeating cycles of 3-4 domains over 10 rounds for adaptation. However, it does not explicitly provide traditional training/test/validation splits for the pretraining dataset or for the target domain data in a manner typically used for supervised learning, as the approach is test-time adaptation where the model continually adapts to incoming data.
Hardware Specification Yes All experiments were conducted on NVIDIA A6000 and RTX 3090 GPUs and further implementation details, including hyper parameters, are in the supplementary material. All runtime and memory measurements were recorded on an NVIDIA RTX 3090 GPU.
Software Dependencies No The paper mentions software components like 'Adam optimizer', 'Co Ex [43]', and 'Mobile Net V2 [44] backbone' but does not specify their version numbers or other key software dependencies like programming languages, frameworks (e.g., PyTorch, TensorFlow), or CUDA versions.
Experiment Setup Yes For the warm-up process, we trained the model using the Adam optimizer for 10 epochs with a fixed learning rate of 5e-4. During test-time adaptation (TTA), we also used Adam optimizer across all methods. The learning rate was set to 5e-6 for training the Adapt BN teacher model and 1e-5 for MADNet2 [45]. Table 7: Learning rates.