Retail Gaze: Gaze Estimation in Retail Environment

RetailGaze consists of 2728 images of the train set, 609 images of the test set, and 585 images of the validation set with 70%, 15%, and 15% as split ratios.
You can download the RetailGaze dataset from the following links:

Google Drive link: https://drive.google.com/file/d/1ZwbloTNhQHX6A_hxAdGHN2Y7iwjcq91x/view?u...
Kaggle link: https://www.kaggle.com/dulanim/retailgaze

The main directory for the images is RetailGaze_V2. The pickle files are,
RetailGaze_V3_2.pickle - The complete annotation set
RetailGaze_V3_2_train.pickle - used for training
RetailGaze_V3_2_test.pickle - used for testing
RetailGaze_V3_2_valid.pickle - used for validation

The contents of the pickle file are a list of dictionary objects with lenght equal to the total number of images. Each object contains the following data:

object = {
    'filename': "the filename tree of this image",
    'width': "640width",
    'height': "480height",
    'gaze_cx': "This is the specific point gx in the image that is considered the gazepoint",
    'gaze_cy': "This is the specific point gy in the image that is considered the gazepoint",
    'ann': {
        'hbox': "Bounding box of the head of the person in the image. In [xmin, ymin, xmax, ymax] format."
    },
    'seg_mask': 'the ground truth segmentation mask tree of the gazed at product area'
}


NOTE: gaze_cx, gaze_yx, hbox all are in the scale of 640x480.