uavaustin / target-finder Goto Github PK

See this for help.

See this link for a starting point. Cache the download of dependencies in the target-finder and target-finder-model builds

• Review images from competition.

• Edit the create_full_images.py script to add some variance (possibly Gaussian distribution) of colors to targets.

Use this reference to add some image sharpening to the output crops of object detection.
2nd reference.

BLOCKED BY #47

Try to remove all the pep8 errors for style

• Add multiprocessing using the the built in python library.

• Refer to create_full_images.py for inspiration.

Create inference pipeline.

https://github.com/dusty-nv/jetson-inference

Currently, color classification is error prone

Investigate the cause for this and potential solutions.

Refer to https://github.com/uavaustin/target-finder/tree/benchmarks for inspiration.

Implement a script that can be used to benchmark our pipeline and return useful analytics such as: classification time, detection time, and accuracy.

Currently we generate data by pasting multiple targets onto a large image, then slice that image up into training data.

This method leads to a problem where targets can be sliced, leading to omitting their existence in training data.

To fix this, create a script which pastes a single target onto slices of the larger background images.

Make skeleton for how targets are displayed on targets page.

Use the GitHub Actions documentation here to create a GitHub Actions workflow for target finder model.

Use this resource for inspiration.

Build a docker image that will be used to optimize a frozen tf graph on the Xavier.

Docker image needs tf < 2.0 and tf/tftrt