Towards In-Field Live Phenotyping for Low-Cost Plant Phenotyping.

We explore how to compress existing very deep CNNs for plant traits segmentation. Thus making them easily deployable in-field and on mobile devices. In particular, we focused on applying these models to the pixel-wise segmentation of plants into multiple classes including the background, a challenging problem in the plant phenotyping community. We combined two approaches (Separable Convolutions and SVD) to reduce model parameters and weight matrices of these very deep CNN-based models. Using our combined method (Separable Convolution and SVD) reduced the weight matrix by up to 95% without affecting pixel-wise accuracy.

Minimum dependencies required to use these codes:

Python 3.6.1

Keras 2.0.6

Tensorflow 1.3.0

Numpy 1.13.3

Pillow 5.1.0

Opencv 3.2.0

Running the codes:

Use train.py to train the baseline models and train_lite.py can be used to train the "Lite" models

Use evaluate.py to evaluate the baseline models and evaluate_lite.py to evaluate the "Lite" models.

Most network parameters can be changed in the parameter file.

All CNN models are contained in the Models folder

Datasets are not included but can be downloaded from:

Flower Dataset: http://www.robots.ox.ac.uk/~vgg/data/flowers/17/index.html

Plant Dataset: https://www.plant-phenotyping.org/datasets-home

CamVid Dataset: http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/