Farm Animal Tracking Project

Project for tracking farm animals.Sample YT

Prerequisites

• Python >=3.7
• TensorFlow Object Detection API

Instalation

Download repository and install dependencies

$ git clone https://github.com/burnpiro/farm-animal-tracking.git
$ cd farm-animal-tracking
$ pip install -r requirements.txt

Download detection model weights

1. To download precompiled model weights Google Drive
2. Unzip archive to model/detection_model

Download recognition model weights

1. To download precompiled model weights Google Drive
2. Unzip archive to model/siamese/weights

Running

Detection

To visualize animal detection on video use:

$ python show_prediction.py

or for image:

$ python run_detection.py

Tracking

To visualize animal tracking on video use:

$ python show_tracking.py --video=<path to video>

Dataset

Dataset for learning of model can be obtained at PSRG website.

Update 2023:

Dataset is no longer available through the university website. Please contact them directly to access the data as described in Issue 8

EDA (Exploratory Data Analysis)

• Run:

docker-compose -f eda/docker-compose.yaml up

• Go to localhost:8001 and enter token from console

Siamese network

You can download current best weights from Google Drive MobileNetV2 Google Drive EfficientNetB5 Google Drive ResNet101V2. Put them into ./model/siamese/weights and use the path as --weights parameter.

Training

Make sure you have cropped dataset in ./data/cropped_animals folder. Please check ./data/data_generator.py documentation for more info.

$ python train_siamese.py

Generate Embeddings for Test dataset and visualize it

Instead of running this script manually (requires ~30GB of RAM) you can use pre-generated train/test/concat files in ./data/visualization. Just select two files with the same postfix, vecs-$1.tsv and meta-$1.tsv, it's important to use the same postfix, otherwise length won't match.

$ python helpers/generate_siamese_emb_space.py

Options:

• --datatype: either train or test (default train), which data should be used for embeddings
• --weights: string (default siam-118_0.0633.h5), specify weights file from mode/siamese/weights/MobileNetV2/ folder

This is going to produce two files:

• vecs.tsv - list of embeddings for test dataset
• meta.tsv - list of labels for embeddings

You can visualize those embeddings in https://projector.tensorflow.org/ application. Just upload them as a custom data (use Load option).

Average class values - Video

Test day data - Video

Train all data - Video

Generate tracking data

$ cd data
$ python generate_tracking.py

This is going to produce tracking data from videos, so we can evaluate model. Look for frames_tracking.json and pigs_tracking.json inside ./data/tracking/. For more details check Wiki.

Testing two images

You can specify the weights for the model. Please use weights marked with the lowest number (loss value).

$ python test_siamese.py

Options:

--weights siam-118_0.0633.h5