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..

DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches

Authors: Yun Xing, Yue Cao, Nhat Chung, Jie M. Zhang, Ivor Tsang, Ming-Ming Cheng, Yang Liu, Lei Ma, Qing Guo

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experimentation, we analyze how variations of this novel structure influence the adversarial effectiveness. We first conduct digital attack experiments with our proposed Depth Vanish patch on KITTI-scene dataset and the four sub-sets of Driving Stereo, i.e., sunny, foggy, rainy, cloudy.
Researcher Affiliation Academia Yun Xing1,2,5 Yue Cao5,6 Nhat Chung5 Jie Zhang5 Ivor Tsang5,6 Ming-Ming Cheng2,3 Yang Liu6 Lei Ma1,4 Qing Guo2,5 1 University of Alberta, Canada 2 VCIP, CS, Nankai University, China 3 NKIARI, Shenzhen Futian, China 4 The University of Tokyo, Japan 5 CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore 6 Nanyang Technological University, Singapore
Pseudocode No The paper describes algorithms such as 'Grid-based Attack' and 'Depth Vanish Attack' in Section 5, detailing the steps and objective functions, but does not present them in a clearly labeled pseudocode or algorithm block.
Open Source Code Yes The code is officially released at: https://github.com/Wi Wi N42/Depth Vanish
Open Datasets Yes For the evaluation of digital attack effectiveness, we adopt the stereo images from KITTI scene flow (KITTI-scene) [23] and Driving Stereo [35] datasets.
Dataset Splits Yes Following [1], 40 stereo image pairs for each (sub-)dataset are selected to verify the effectiveness of different patches.
Hardware Specification No The paper refers to Intel Real Sense D435i depth camera for physical evaluation, but does not provide specific hardware details (e.g., GPU/CPU models, memory) used for optimizing or training the adversarial patches within the main text. The justification for NeurIPS checklist states that this information is in the supplemental materials, but not in the main paper.
Software Dependencies No The paper does not explicitly state specific versions for software dependencies such as libraries or programming languages in the main text. It mentions general components like 'DNN-based stereo pipelines' but lacks detailed version numbers.
Experiment Setup Yes During the optimization for the Grid-based and Depth Vanish attacks, we use a patch with a physical size (u, v) = (0.891 m, 1.26 m) and specify the physical ground-truth depth e = 5 m. To assemble the texture element into a patch, we empirically set the number of repetition as 5 for horizontal and 4 for vertical, i.e., k = (4, 5). For the optimization and corresponding evaluation results with different patch physical setup, we provide them in the supplemental material. When the patch is optimized as grid-based attack, the optimal interval size o = 10 px from Sec. 3 is applied. As for the loss weights adopted during the depth vanish attack, we keep setting α = 0.1 and β = 10.