This repository contains scripts to help process the Vehicle Rear Signal Dataset provided by UC Merced.

Notebooks

process_rear_signal_dataset.ipynb

This notebook will extract a single frame from every sequence to use for training. This is useful for creating datasets for one-shot models.

Output Dataset folder structure: One image from every sequence is moved into a directory corresponding to that sequence's label

|_ BLO
|  |_ frame_000001.png
|  |_ frame_000002.png
|  |_ ...
|_ BLR
|  |_ frame_000001.png
|  |_ frame_000002.png
|  |_ ...
|_ ...

process_rear_signal_dataset_sequences.ipynb

This notebook is used to organize the data in such a way that it can be used for training models that handle sequences of data (LSTM, SlowFast, ResNet3D, etc).

Output Dataset folder structure:

|_ frames
|  |_ [video name 0]
|  |  |_ frame_000001.jpg
|  |  |_ frame_000002.jpg
|  |  |_ ...
|  |_ [video name 1]
|     |_ frame_000001.jpg
|     |_ frame_000002.jpg
|     |_ ...
|_ train.csv
|_ test.csv
|_ val.csv
|_ label_to_id.pickle (maps from label string -> class id #)

Dataset Description

This dataset contains time-sequence images of different vehicle rears under various real-world road conditions. Each image frame of the sequence is the vehicle rear cropped mannually from the raw video. Every sequence is categorized into 8 classes regarding to the status of the rear lights (brake and turn lights).

Class definition

This data set includes sequences labeled with a total of 8 distinct classes for all possible rear signal states. Each state is denoted by 3 letters: B (brake), L (left), and R (right). We give either the corresponding letter of the signal when it is on, or a letter O for off. Consequently, there will be 8 different states: OOO: Brake light and turn signals off BOO: Brake light on, turn signals off OLO: Brake light off, left signal on BLO: Brake light on, left signal on OOR: Brake light off, right signal on BOR: Brake light on, right signal on OLR: Brake light off, left and right signal on (hazard warning light on) BLR: Brake light on, left and right signal on (hazard warning light on)

Folder structure

-rear_signal_dataset -Footage_name -Footage_name_XXX -Footage_name_XXX_DDD (sequence of class XXX starting from frame number DDD) -light_mask -frameDDDDDDDD.png (frames with a 8 digit number indicating the frame number) -... -Footage_name_XXX_DDD -light_mask -frameDDDDDDDD.png -... -Footage_name_XXX -Footage_name_XXX_DDD -... -Footage_name_XXX_DDD -... -Footage_name_XXX_DDD -... -Footage_name -... *XXX indicates the label of the signal

Dataset statistics

Total sequences: 649 Total frames: 63637

Number of sequences in each class: OOO: 188 BOO: 211 OLO: 78 BLO: 63 OOR: 58 BOR: 33 OLR: 9 BLR: 9

Number of frames in each class: OOO: 21867 BOO: 17874 OLO: 6271 BLO: 6380 OOR: 4728 BOR: 3527 OLR: 1600 BLR: 1390

Level of difficulty

There is three levels of difficulty for this dataset: Easy, Moderate and Hard. There are three .txt files repectively documenting the sequence folder names(e.g. 20160805_g1k17-08-05-2016_15-57-59_idx99_BOO_00002671) in that specific level.

EricWiener/rear_signal_dataset_processing