https://github.com/agentmorris/und-ducks
Code for training and evaluating a detector for the UND ducks dataset
https://github.com/agentmorris/und-ducks
Last synced: about 1 year ago
JSON representation
Code for training and evaluating a detector for the UND ducks dataset
- Host: GitHub
- URL: https://github.com/agentmorris/und-ducks
- Owner: agentmorris
- License: mit
- Created: 2023-06-18T22:58:01.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2023-09-23T01:44:39.000Z (over 2 years ago)
- Last Synced: 2025-02-10T18:11:18.231Z (over 1 year ago)
- Language: Python
- Size: 297 KB
- Stars: 2
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
## UND duck detector training
### Overview
The code in this repo trains and evaluates models to detect ducks and other waterfowl in drone images, based on a dataset provided by the [Felege Wildlife Lab](https://arts-sciences.und.edu/academics/biology/susan-felege/).
There are around 35,000 images total, about 95% of which contain no birds. Images are around 5000x3000. A typical ground truth image looks like this:
The annotations you can vaguely see as different colors correspond to different species of bird. Most of this repo operates on 1280x1280 patches that look like this (this is also a ground truth patch, not a model prediction):
### Files
These are listed in roughly the order in which you would use them.
#### und-ducks-data-import.py
* Match images to annotation files
* Read the original annotations in Pascal VOC format
* Convert to COCO format
* Do a bunch of miscellaneous consistency checking
#### und-ducks-training-data-prep.py
* For all the images with at least one annotation, slice into mostly-non-overlapping patches
* Optionally sample hard negatives (I did not end up actually using any hard negatives)
* Split into train/val
* Export to YOLO annotation format
#### und-ducks-training.py
* Train the model (training happens at the YOLOv5 CLI, but this script documents the commands)
* Run the YOLOv5 validation scripts
* Convert YOLOv5 val results to MD .json format
* Example code to use the MD visualization pipeline to visualize results
* Example code to use the MD inference pipeline to run the trained model
### Running the models, and doing stuff with the results
Much of the code in this repo is shared with the [Izembek Brant Goose Detector](https://github.com/agentmorris/usgs-geese) repo; the training code is bespoke for each model, but the inference code is identical. So, for instructions on running the model and postprocessing the results, see the "[Running the model](https://github.com/agentmorris/usgs-geese#running-the-model)" section on the Izembek Detector repo.