Toward a Taxonomy and Computational Models of Abnormalities in Images
Authors: Babak Saleh, Ahmed Elgammal, Jacob Feldman, Ali Farhadi
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We propose a new dataset of abnormal images showing a wide range of atypicalities. We design human subject experiments to discover a coarse taxonomy of the reasons for abnormality. Our experiments reveal three major categories of abnormality: object-centric, scene-centric, and contextual. Based on this taxonomy, we propose a comprehensive computational model that can predict all different types of abnormality in images and outperform prior arts in abnormality recognition. |
| Researcher Affiliation | Collaboration | Babak Saleh Dept. of Computer Science Rutgers University New Jersey , USA Ahmed Elgammal Dept. of Computer Science Rutgers University New Jersey , USA Jacob Feldman Center for Cognitive Science, Dept. of Psychology Rutgers University New Jersey , USA Ali Farhadi Allen Institute for AI, & Dept. of Computer Science University of Washington Washington , USA |
| Pseudocode | No | No pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | No | With the final version of this paper we will publish the dataset along with the human subject experiment results and implementation of our computational models. |
| Open Datasets | Yes | With the final version of this paper we will publish the dataset along with the human subject experiment results and implementation of our computational models. |
| Dataset Splits | No | We perform fivefold cross validation to find the best values for parameters of the SVM. This achieves in 87.46% average precision for the task of object classification in PASCAL2010 test set. Numbers in parenthesis show the reported errors on normal images (ILSVRC 2012 validation data), while numbers next to them is the error on our abnormal images. |
| Hardware Specification | No | No specific hardware (GPU, CPU models, etc.) is mentioned in the paper for running experiments. |
| Software Dependencies | No | The paper mentions software components like 'Kernel Descriptors', 'SVM classifiers', 'HOG', 'color SIFT', and 'Caffe', but no specific version numbers are provided for any of them. |
| Experiment Setup | Yes | We specifically use Gradient Match Kernels, Color Match Kernel, Local Binary Pattern Match kernels. We compute these kernel descriptors on fixed size 16 x 16 local image patches, sampled densely over a grid with step size 8 in a spatial pyramid setting with four layers. This results in a 4000 dimensional feature vector. We train a set of one-vs-all SVM classifiers for each object class using normal images in PASCAL train set. We perform fivefold cross validation to find the best values for parameters of the SVM. |