Learning to count leaves

Torch7 code for training a CNN for semantic and instance segmentation of leaves.

Directory structure

Directory	Content
CNN	Core CNN code
data	Training data
shells	various bash shells

Dependencies

install dependencies

> shells/dependencies.sh

Data

We used Plant Phenotyping DataSet for training. See Plant-phenotyping and how it was collected. Download the dataset and copy to

> $ROOT/data

see dataset CLI argument for adjusting the path to desired location.

Training

See opts.lua for CLI options.

> th -i main.lua

To start anew:

> rm -rf $ROOT/CNN/results/
> rm $ROOT/data/NLL_trainData.t7 and/or rm $ROOT/data/BCE_trainData.t7
> th main.lua

To generate data only:

> rm -rf $ROOT/CNN/results/
> rm $ROOT/data/NLL_trainData.t7 and/or rm $ROOT/data/BCE_trainData.t7
> th main.lua -genDataOnly

To see datasets:

> rm -rf $ROOT/CNN/results/
> rm $ROOT/data/NLL_trainData.t7 and/or rm $ROOT/data/BCE_trainData.t7
> qlua main.lua 

Inference

See inference.lua for CLI options.

> th -i inference.lua

Related (CNN-based) Efforts

• Recurrent Instance Segmentation (Does Leaf counting)

  • code for RIS

  • Paper that inspired RIS

• DEEP-PLANT: PLANT IDENTIFICATION WITH CONVOLUTIONAL NEURAL NETWORKS

• Plant recognition using CNNs

• Fine-tuning Deep Convolutional Networks for Plant Recognition

• Resnet training in Torch

• Faster_rcnn

• Instance-aware Semantic Segmentation via Multi-task Network Cascades

• Deepmask

• Mask RCNN

• Berkely Fully Convolutional Networks

• Leafsnap: A Computer Vision System for Automatic Plant Species Identification code

TODO

• patch-based
• post process with CRF