SSD300

Tutorials

• Introduction to code.
• Dataset setup (Not required for TDT4265 computers).
• Running tensorboard to visualize graphs.

Install

Follow the installation instructions from previous assignments. Then, install specific packages with

pip install -r requirements.txt

Dataset exploration

We have provided some boilerplate code for getting you started with dataset exploration. It can be found in dataset_exploration/analyze_stuff.py. We recommend making multiple copies of this file for different parts of your data exploration.

To run the script, do the following command from the SSD folder:

python -m dataset_exploration.analyze_stuff

Dataset visualization

We have also created a script visualizing images with annotations. To run the script, do

python -m dataset_exploration.save_images_with_annotations

By default, the script will print the 500 first train images in the dataset, but it is possible to change this by changing the parameters in the main function in the script.

Qualitative performance assessment

To check how the model is performing on real images, check out the performance assessment folder. Run the test script by doing:

python -m performance_assessment.save_comparison_images <config_file>

If you for example want to use the config file configs/tdt4265.py, the command becomes:

python -m performance_assessment.save_comparison_images configs/tdt4265.py

This script comes with several extra flags. If you for example want to check the output on the 500 first train images, you can run:

python -m performance_assessment.save_comparison_images configs/tdt4265.py --train -n 1000

Test on video:

You can run your code on video with the following script:

python -m performance_assessment.demo_video configs/tdt4265.py input_path output_path

Example:

python3 -m performance_assessment.demo_video configs/tdt4265.py Video00010_combined.avi output.avi

You can download the validation videos from OneDrive. These are the videos that are used in the current TDT4265 validation dataset.

Bencharking the data loader

The file benchmark_data_loading.py will automatically load your training dataset and benchmark how fast it is. At the end, it will print out the number of images per second.

python benchmark_data_loading.py configs/tdt4265.py

Uploading results to the leaderboard:

Run the file:

python save_validation_results.py configs/tdt4265.py results.json

Remember to change the configuration file to the correct config. The script will save a .json file to the second argument (results.json in this case), which you can upload to the leaderboard server.