beehive: counting bees using AI.

We want to count the number of bees flying around in order to monitor the hive. The goal of this challenge is to automate the process of counting bees in a given image.

beehive uses a slightly modified Centernet to achieve this.

Solution Documentation

For methodolgy and results, check report.md or report.pdf

Get Started

clone project: git clone https://github.com/aadhithya/beehive.git

Inference

bee_counter.py is a minimal inference script that runs inference using onnxruntime.
Insatll requirements: pip install -r inference_requirements.txt
run inference: python bee_counter.py <image-path> --show True
NOTE: The checkpoint is downloaded from github if not available locally. Check Releases for checkpoints.

Training and Development

In case you want to develop or train/evaluate/infer model you need to do the following:

Set up using poetry

install poetry: pip install poetry
install requirements: poetry install

Note: poetry install creates a new virtual env.
now that the environment is created, you are ready to go.
check installation: python -m beehive version

NOTE: using this method needs python >= 3.9

(or)

Set up using pip

install requirements: pip install -r requirements.txt
install beehive in edit mode: pip install -e .
check installation: python -m beehive version

Download pre-trained weights

When you inference for the first time, the corresponding model weights are automatically downloaded to the current directory from github. : When using python -m beehive infer ... you need to specify the --dl flag to enable model download.
Models can be manually downloaded from the Releases.
Only weights for the resnet18 backbone model is available.

Train Model

python -m beehive train --help

The splits.json file contains the dataset splits.

Evaluate Model

python -m beehive eval --help

pytorch Inference

python -m beehive infer path/to/image --show

ONNX Export

python -m beehive export-onnx --help

aadhithya/beehive