The summary report is found under "cv_assignment_2_report.pdf".
The primary parquet file identified in the deliverable is the "all_detections.parquet" file. This file contains all detections within the video, which was sampled at one frame for every five seconds of video. The schema for this file is as follows: video_id - the URL of the YouTube video on which the segmentation was performed. source_file - the name of the file under which the individual frame was saved (the number after the underscore is the frame index) name - the class label as a string class - the class label as an integer confidence - the confidence score box - the bounding box (x_min, y_min, x_max, y_max) segments - the detection masks
Contiguous detections of each class can be found in the "contiguous_class_detection.parquet" file. The schema for this file is as follows: start_timestamp - the starting timestamp when the class is identified within the video during a contiguous period end_timestamp - the ending timestamp when the class is identified within the video during a contiguous period start_frame_no - the frame number corresponding to the start_timestamp end_frame_no - the frame number corresponding to the end_timestamp name - the class label as a string class - the class label as an integer number_of_supporting_detections - the number of detections of the class between the start and end timestamps
An example of the semantic image search, which takes as an input an image and outputs the contiguous detections of the car parts contained within the image, can be found in the "contiguous_search_output_example.parquet" file. The schema of this file is the same as the "contiguous_class_detection.parquet" file schema described above.