This dataset contains the images and labels used in this paper: [Unsupervised Outlier Detection in Appearance-Based Gaze Estimation](https://openaccess.thecvf.com/content_ICCVW_2019/papers/RLQ/Chen_Unsupervised_Outlier_Detection_in_Appearance-Based_Gaze_Estimation_ICCVW_2019_paper.pdf).

Data structure:
```
	___p_01
	    |___1_7.mat
	    |___1_8.mat
	    |___1_9.mat
	___p_02
	    |___2_7.mat
	    |___2_8.mat
	    |___2_9.mat
	...
```
Each folder named `p_{subject_id}` contains three `.mat` files. Each `.mat` file contains a sequence of images shaped as `(N, C, W, H)`. The naming convention is `{subject_id}_{location_id}.mat`.

The `ICCVW_labels.csv` contains the categorical label for each image, formated as below:
`|subject_id | location | image_idx | label|`
where `label` is one out of five categories: `['Normal', 'Blink', 'Occlusion', 'Blur', 'Landmark failure']`.

The number of images in each `{subject_id}_{location_id}.mat` are listed here:
```
subject_id = 1, idx_offset = [2956 2957 2947]
subject_id = 2, idx_offset = [2949 2956 2789]
subject_id = 3, idx_offset = [2967 2955 2962]
subject_id = 4, idx_offset = [2931 2952 2947]
subject_id = 5, idx_offset = [2989 2990 2984]
subject_id = 6, idx_offset = [2989 2984 2988]
subject_id = 7, idx_offset = [2939 2998 2959]
subject_id = 8, idx_offset = [2971 2993 2979]
subject_id = 9, idx_offset = [2989 2976 2961]
subject_id = 10, idx_offset = [2986 3027 2993]
subject_id = 11, idx_offset = [2992 2990 2989]
subject_id = 12, idx_offset = [2956 2953 2962]
subject_id = 13, idx_offset = [2941 2938 2799]
subject_id = 14, idx_offset = [2910 2760 2874]
subject_id = 15, idx_offset = [2942 2741 2826]
subject_id = 16, idx_offset = [2937 2844 2910]
subject_id = 17, idx_offset = [2864 2887 2857]
subject_id = 18, idx_offset = [2957 2956 2960]
subject_id = 19, idx_offset = [2959 2927 2952]
subject_id = 20, idx_offset = [2919 2887 2919]
subject_id = 21, idx_offset = [3018 3015 3053]
```
`idx_offset = [N_7, N_8, N_9]` is a list containing the number of images in location 7, 8 and 9.