This is an animal detection project based off of PyTorch Tutorial to Object Detection.

The dataset used for training and validation is accessible on https://lila.science/datasets/snapshot-serengeti. The specific images used are listed in the bbox_images_split.csv file in the snapshot-serengeti folder.

The conda environment used to run the program can be setup using the environment.yaml file, using the command conda env create -f environment.yaml, and activated with conda activate environment.

Executing Sample Code

Evaluation of a fully trained model (on a small sample of the validation dataset) can be performed by:

1. Activating the conda virtual environment on a machine with CUDA enabled
2. Change into the repository source directory
3. Run the command python eval.py.

To run object detection on a single image:

1. Activate the conda virtual environment
2. Change into the repository source directory
3. Run the command python detect.py.
4. The image will be output to sample.png for inspection.

Project Structures

Datasets

datasets.py: contains the class used to retrieve and handle the bounding box dataset snapshot-serengeti/: contains dataset metadata and a sample of the images in the dataset

Transformations

transformations.py: code with the custom image and bounding box transformations, as well as train/validation sequential transforms

Utils

utils.py: miscelaneous utility functions

Eval

eval.py: functions for evaluating the mAP of various models.

Train

train.py: the main training code for the model

Models

models/EfficientNetSSD300.py: the SSD300 model with EfficientNetB2 base feature extractor. This was the final model used for training. models/EfficientNetSSD300.py: the backup SSD300 model using MobileNet as the base feature extractor.