donfaq / cnn-rnn

CNN-RNN for human activity recognition

CNN-RNN

Tensorflow based implementation of convolution-reccurent network for classification of human interactions on video.
Uses SDHA 2010 High-level Human Interaction Recognition Challenge dataset.

Requirements

• Python 3.5.2
• Tensorflow > 1.0
• Python OpenCV > 3.0

Evaluation

Example:

python main.py --lrate 0.001 --update

Parameter	Default value	Description
epoch	1	Number of epoch
esize	50	Size of examples
estep	20	Length of step for grouping frames into examples
height	240	Height of frames
width	320	Width of frames
lrate	1e-4	Learning rate
logdir	network/logs	Path to store logs and checkpoints
conv	standard	Type of CNN block (inception/vgg16)
rnn	GRU	Type of RNN block (LSTM/GRU)
update	False	Re-Generate TFRecords
download	False	Download dataset
restore	False	Restore from previous checkpoint
test	False	Test evaluation