Adaptive Hierarchical Certification for Segmentation using Randomized Smoothing
Authors: Alaa Anani, Tobias Lorenz, Bernt Schiele, Mario Fritz
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experiments on the datasets Cityscapes, PASCAL-Context, ACDC and COCO-Stuff demonstrate that our adaptive algorithm achieves a higher CIG and lower abstain rate compared to the current state-of-the-art certification method. |
| Researcher Affiliation | Academia | 1CISPA Helmhotz Center for Information Security, Saarbr ucken, Germany 2Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbr ucken, Germany. |
| Pseudocode | Yes | These steps are outlined in Algorithm 1 describing GETCOMPONENTLEVELS... Algorithm 2 HSAMPLE... Algorithm 3 ADAPTIVECERTIFY... Algorithm 4 SAMPLEPOSTERIORS... Algorithm 5 HYPOTHESESTESTING. |
| Open Source Code | Yes | Our code can be found here: https://github.com/Alaa Anani/adaptive-certify. |
| Open Datasets | Yes | We use four segmentation datasets: Cityscapes (Cordts et al., 2016), the Adverse Conditions Dataset with Correspondences (ACDC) (Sakaridis et al., 2021), PASCAL-Context (Mottaghi et al., 2014) and COCO-Stuff-10K (Caesar et al., 2018) |
| Dataset Splits | Yes | The final model performance on the clean split has a mean per-pixel accuracy of 62.77% and an m Io U of 0.3146. Meanwhile, on the noisy validation split, the mean per-pixel accuracy is 53.43% and the m Io U is 0.2436. We validate twice every epoch on both the clean and noisy (σ = 0.25) validation splits. |
| Hardware Specification | No | The paper mentions training was done 'per GPU' but does not specify any particular GPU models, CPU models, memory amounts, or other specific hardware configurations used for the experiments. |
| Software Dependencies | No | The paper mentions 'PyTorch' and 'Paddle Cls' as software used, but it does not provide specific version numbers for these or any other ancillary software components, which is required for reproducibility. |
| Experiment Setup | Yes | Unless stated otherwise, all certification results use the values σ = 0.25, τ = 0.75 and α = 0.001 in both our algorithm ADAPTIVECERTIFY and the baseline SEGCERTIFY. We use different parameters for the threshold function Tthresh (Eq 4) for ADAPTIVECERTIFY per dataset as described in App. B.3, which we found via a grid search that maximizes the CIG metric. The batch sizes used for training and validation were 12 and 1 respectively, per GPU. |