Detection as Regression: Certified Object Detection with Median Smoothing

Authors: Ping-yeh Chiang, Michael Curry, Ahmed Abdelkader, Aounon Kumar, John Dickerson, Tom Goldstein

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We use the MS-COCO dataset [20] to test the resulting detector, obtaining the first detector to achieve non-trivial ℓ2-norm certified average precision on large scale image dataset. We mainly use YOLOv3 [28], pretrained on the MS-COCO dataset [20], as our black-box detector where Io U thresholds for NMS is set to 0.4. The evaluation is done on all 5000 images from the test set, with adversarial perturbations δ 2 < ϵ = 0.36.
Researcher Affiliation Academia Ping-yeh Chiang University of Maryland pchiang@cs.umd.edu Michael J. Curry University of Maryland curry@cs.umd.edu Ahmed Abdelkader University of Maryland akader@cs.umd.edu Aounon Kumar University of Maryland aounon@cs.umd.edu John Dickerson University of Maryland john@cs.umd.edu Tom Goldstein University of Maryland tomg@cs.umd.edu
Pseudocode Yes Algorithm 1 Prediction and Certified Detection
Open Source Code No The paper does not provide a specific repository link, explicit code release statement, or code in supplementary materials for the methodology described in this paper.
Open Datasets Yes We use the MS-COCO dataset [20] to test the resulting detector
Dataset Splits No The paper mentions using a 'test set' but does not provide specific train/validation/test split information such as exact percentages, sample counts for each split, or citations to predefined splits.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions software components like YOLOv3, Mask RCNN, Faster RCNN, and DNCNN denoiser, but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Io U thresholds for NMS is set to 0.4. To perform the smoothing, we inject Gaussian noise with standard deviation σ = 0.25, and we use 2000 noise samples for each image. The estimated upper and lower bound for each coordinate are selected such that they bound the true hp(x) and hp(x) with confidence α = 99.999%.