metrics

experiments/metrics

Code for computing metrics over predictions.

• aggregate_metrics.py script computing the mean of metrics in different sequences. This relies on txt files precomputed through compute_metrics.py

• aggregate_metrics_by_scenario.py script that merges a metric in different scenarios, and prints a mean for all couples (time_of_day, wheather)

• calc_mean_gt_for_new_gt.py script that finds the mean groundtruth fixation map, which is considered a baseline.

• compute_metrics.py script for computing metrics over predictions of a sequence and save them to txt files.

Usage: python compute_metrics.py --start <start_sequence> --stop <stop_sequence>

TODO: maybe merge the three different methods in one?

• compute_metrics_on_attentive.py script to compute metrics on attentive subsequences only. This relies on txt files precomputed through compute_metrics.py

• eval_semseg_by_sequence.py script that proposes random samples of semantic segmentation to the user that has to decide if is good or bad pressing 'g' or 'b' respectively. Selected sequences (where segmentation is good) are used to test the semantic segmentation performance on an additional ablation study where terrible segmentations are excluded.

• segmentation_stats.py is a script that reports scenario and actions distributions in frames in which the segmentation branch works better than other ones.