Axiomatic Explanations for Visual Search, Retrieval, and Similarity Learning
Authors: Mark Hamilton, Scott Lundberg, Stephanie Fu, Lei Zhang, William T. Freeman
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 EXPERIMENTAL EVALUATIONResults In Table 1 and Table 4 of the Supplement we report experimental results for Pascal VOC and MSCo Co respectively. |
| Researcher Affiliation | Collaboration | Mark Hamilton1,2, Scott Lundberg2, Stephanie Fu1, Lei Zhang2, William T. Freeman1,3 1MIT, 2Microsoft, 3Google |
| Pseudocode | No | The paper does not include a figure, block, or section labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | We also provide training and evaluation code at https://aka.ms/axiomatic-code |
| Open Datasets | Yes | We evaluate our methods on the Pascal VOC (Everingham et al., 2010) and MSCo Co Caesar et al. (2018) semantic segmentation datasets. |
| Dataset Splits | Yes | Data: For evaluations within Table 1 we use the Pascal VOC (Everingham et al., 2010) dataset. In particular we form a paired image dataset by using Mo Co V2 to featurize the training and validation sets. All experiments use images that have been bi-linearly resized to 224 224 pixels. For each image in the Pascal VOC validation set we choose a random segmentation class that contains over 5% of image pixels. We then find each validation image s closest Conditional Nearest Neighbor (Hamilton et al., 2020) from the images of the training set of the chosen segmentation class. |
| Hardware Specification | Yes | Experiments use Py Torch (Paszke et al., 2019) v1.7 pre-trained models, on an Ubuntu 16.04 Azure NV24 Virtual Machine with Python 3.6. |
| Software Dependencies | Yes | Experiments use Py Torch (Paszke et al., 2019) v1.7 pre-trained models, on an Ubuntu 16.04 Azure NV24 Virtual Machine with Python 3.6. |
| Experiment Setup | Yes | For sampling procedures such as LIME, Kernel SHAP, and Partition SHAP we use 5000 function evaluations. For first and second-order super-pixel based methods (LIME, Kernel-SHAP) we use the SLIC superpixel method (Achanta et al., 2010) provided in the Scipy library (Virtanen et al., 2020) with 50 segments, compactness = 10, and σ = 3. For SBSM we use a window size of 20 pixels and a stride of 3 pixels. We batch function evaluations with minibatch size 64 for backbone networks and 64 20 for SAM based methods. For all background distributions we blur the images with a 25-pixel blur kernel with the exception of LIME and SBSM which use mean color backgrounds. For second order methods we use the same background and superpixel algorithms, but implement all methods within Py Torch for uniform comparison. For SBSM, Kernel SHAP, and LIME we use 20000 samples and for KSAM and IGSAM we use 40000 samples. |