Camera Trap video processing for Manacus dynamics assessment
- Dataset Preparation
- Extract video frames given the metadata sheet using ./dataset/fcat/locate_dataset.py
- Organize the dataset for bird detection
Stage 1
: ebird high resolution images in coco format ./dataset/DATASET.mdStage 2
: camera trap image sampling [TODO]
- Optical flow or Background subtraction to determine the most likely frames to contain birds for detection
- Run the videos through Marco's audio processing pipeline for likely frames (eg: Bird snaps right after flying off or copulation)
- Manacus detection
- Detect bird/manacus in the video frames -
bird
,background
[Deprecated] -
Stage 1
model to detect manacus in image frames (male/female should be present)- male (white chest, black wings) [only male displays aand makes snap noise]
- female (green) [low representation, camouflages with the foliage]
- unknown (juvenile: between male and female) [male bird can be practicing infront of juveniles ]
-
Stage 2
model trained via transfer learning on camera trap manacus dataset
- Detect bird/manacus in the video frames -
- Dynamics
- Estimate copulation/visitation in the video based on heuristics
- frame by frame position (bounding box) for male and female
- Estimate copulation/visitation in the video based on heuristics
- Python 3.8.18 Installation
- Instructions via Miniconda v23.1.0 - https://docs.conda.io/projects/miniconda/en/latest/
- Create a virtual environment named "mothra" for this analysis
conda env remove -n mana conda create -n mana python=3.9 jupyterlab -c conda-forge conda activate mana pip install -r requirements.txt
- Clone the current codebase -
git clone https://github.com/karmarv/manacus-dynamics.git && cd manacus-dynamics
- Install pre-requisite packages in the activated python virtual environment using -
pip install -r requirements.txt
- Refer ./dataset/DATASET.md
- Refer ./model/MODEL.md