Given an input video, you can use this to output a log file of bounding boxes around traffic signs/lights, cars, bicycles, and motorcycles. Each line of the log file is in the format:
frame # | x_min | x_max | y_min | y_max
You can also use a folder of images as input, in case there are issues with converting the list of images into a video.
Python 3 is required; the repository has been tested with Python 3.6 running on Ubuntu 16.04 and 18.04.
Some of the files in this repository are stored using Git Large File Storage. It is easiest to set up Git LFS before cloning this repository. Otherwise, use git lfs pull after cloning and initializing Git LFS to ensure you have the full trained models.
If using pip to manage dependencies:
- pip install opencv-python absl-py matplotlib numpy pillow
- pip install tensorflow-gpu==1.14
Run the following from the root directory
python detection.py --model_path=full_trained_model/detectors-40264 --video_path=/path/to/video/file --min_threshold=0.70 --output_path=/path/to/output/folder
Threshold determines the level of certainty required for a bounding box to be reported (higher values result in more false positives).
To use a different model, you can change model_path accordingly (ex. --model_path=/trained_model/detectors-9614).
The detection log is saved as {video_name}_log.txt (ex. scenario1_log.txt for scenario1.avi).
Run the following from the root directory
python image_detection.py --model_path=full_trained_model/detectors-40264 --images_path=/path/to/images/folder --min_threshold=0.70 --output_path=/path/to/output/folder
Threshold and model_path are the same parameters as described above.
The detection log is saved as {folder_name}_log.txt (ex. rgb_log.txt for scenario1/rgb/).
Add a --save_images argument to either of the prior scripts to output the frames (.png) and a video concatenating the frames (.avi). If running the script on a folder of images, the frames will be concatenated in the same order as the sorted file names.
If you want to convert a folder of images to a video instead, you can run:
python image_to_video_converter.py --image_folder=/path/to/image/folder --video_path=/path/to/output/video --fps=desired_integer_fps
The log file storing the mAP scores (and some other metrics) is located at: Evaluation Metrics. The results are in ascending order of the number of iterations for the models present in trained models
For a cleaner representation, consider exporting the csv from tensorboard as described below.
If you can view tensorboard results on your machine, use the following from the root directory:
tensorboard --logdir evaluations/full_evaluation
You can scroll to the bottom of the rendered page and either save the mAP plot image as a .svg or export the mAP values as a .csv
You can turn smoothing to 0 to get an exact plot
Similarly to visualizing the mAP metric, you can use the following from the root directory to visualize training loss metrics:
tensorboard --logdir full_trained_model/checkpoints
The model was trained and evaluated with the data provided in: https://github.com/DanielHfnr/Carla-Object-Detection-Dataset