IAC_dataset_maker

A simple repository with a pipeline to identify and extract important camera data to be labelled.

Dataset Pipeline

Use the following steps to extract data from a rosbag, remove image distortions, choose image segments with vehicles, and create a dataset to be labelled:

1. multi_extract.sh: This is a bash script to extract images from a rosbag and optionally undistort them during extraction. To set his up for your own rosbag follow the following steps:

   1. Specify the extraction environment variables:
      • DATA_DIR_DEFAULT: Directory where rosbags exist. We will search this directory for db3 files to determine where these rosbags are.
      • VERBOSE_DEFAULT: Set to 1 to get verbose extraction.
      • UNDISTORT_DEFAULT: Set to 1 to get undistortion during the extraction process.
      • CALIB_DIR_DEFAULT: If you choose to undistort, pass the path to the directory with the calibration files.
   2. [OPTIONAL] Instead of specifying the DEFAULT VARIABLES as in step 1, you may export environment variables in the bash shell before running. Here is an example:

      •   export DATA_DIR=<PATH_TO_ROSBAGS_DIR>
          export VERBOSE=1
          export UNDISTORT=1
          export CALIB_DIR_DEFAULT=<PATH_TO_CALIBRATION_FILES>

   3. Make sure the topic names you want to extract exist in the ros2bag_image_extractor.py in the iterator dictionary.
   4. Run multi_extract.sh:

          bash multi_extract.sh

   5. Now sit back and wait until the extraction is complete. This process can take and hour or longer depending on the size of the rosbag.
   6. [IMPORTANT PLEASE READ] By setting the variable MAKE_VID_DEFAULT in the step 1.1 or by export it as an environment variable in step 1.2, you can skip step 2 in the first pass.

2. This step can be accomplished by setting Convert the extracted image data to a video to view the segments that contain the car. This is important to reduce the dataset size that is sent to an external vendor for labelling, saving cost and time. Use the following command in the OUTPUT_DIR printed by the previous command:

   ffmpeg -framerate 50 -pattern_type glob -i '*.jpg' -c:v libx264 -profile:v high -crf 20 -pix_fmt yuv420p <NAME_AND_PATH_OF_VIDEO_FILE_OUTPUT.mp4>

3. [COMPLETELY MANUAL] Review the video files and note down the timestamps where the vehicle starts being visible and where it stops being visible. Convert the timestamp into seconds format and multiply by 50 (The frame rate we set in the previous step) to get the sequence number for each start and stop position.

   • For example a time of 1:10 (1 minute and 10 seconds) in the video viewer translates to 70 seconds and a frame ID of 3500 (70*frame_rate, where frame_rate is 50 for our example)

4. dataset_maker.py: Next we we will extract these important segments into a seperate folder for final extraction and dataset creation.

   • Input the sequence numbers from the previous step into the dataset_maker.py file in the ranges array in pairs, for example:

     ranges = np.array(
         [
             0, 3500,
             30000, 43500,
             66000, 74250,
             86500, 87750,
         ]
     )

   • Set the source directory and destination directory in the dataset_maker.py file:

         SOURCE_DIR = "<ONE OF THE EXTRACTED DATA DIRS FROM STEP 1>"
         DEST_DIR = "<PATH WHERE DATASET IS TO BE OUTPUT>"

   • Now run the python file

         python3 dataset_maker.py
     

5. Repeat Steps 2 and 3 for the extracted segments of images, to make sure that you have not removed images that had the car but you stopped early, or images that had the car and you started late in your ranges. If you find such discrepancy go back and adjust ranges appropriately and rerun step 4. Repeat this step (step 5), until you no longer make a mistake.

HELPFUL TIPS

1. Try to get ranges that give atleast a 5-10 second margin from when the car enters and when it exits.
2. Watch the videos really carefully, and do not skip (its ok to go fast, but do not skip video frames) as there are instances when a car may only be available for a a second or two.

Contact

• Aman Saraf | [email protected]
• Chris Lai | [email protected]