Action Recognition NN

This project is a set of example workarounds for action recognition & deep learning. It's based on PyTorch and aims to classify given still images for human actions.

Dataset

The dataset that this project is based is the Stanford 40 Actions, containing more than 9.500 images, capturing human actions. To download the dataset check here

Design

The project is based on two main packages, models and sfd40:

• models: Covers all the functionality around the neural networks.
• sfd40: Covers all the functionality around the Stanford 40 data load.

Both projects have a manager class which is the main class for every package. Those two classes play important role in the main.py file.

Installation

The default approach shown in the readme is based on uv, however one can install all dependencies using other tools as well. To install uv check here.

Set Dataset Paths

First, we start by exporting the xml and image directories so our script is able to fetch the two different dirs:

export IMAGE_FILES_PATH="absolute-path-stanford40-jpeg-images-dir"
export XML_FILES_PATH="absolute-path-stanford40-xml-annotations-dir"

Usage

In order to run the Neural Network example you can simply run:

make run

Available Configurations

The script can be configured through the usage of environment variables. An example usage with custom configuration is:

# Increased number of epochs
NN_NUM_EPOCHS=500 make run

# Increased number of epochs and only pretrained model selected
NN_NUM_EPOCHS=500 MODEL="pretrained" make run

The env vars used are:

Data Ratio

Name	Description	Type	Default
VALIDATION_RATIO	The percentage of train data used for validation	float	0.05
TEST_RATIO	The percentage of full data used for test	float	0.15

Directory Paths

Name	Description	Type	Default
IMAGE_FILES_PATH	The path to the jpeg images dir	string	"JPEGImages"
XML_FILES_PATH	The path to the xml annotations dir	string	"XMLAnnotations"

Neural Network HyperParameters

Name	Description	Type	Default
NN_IMAGE_READ_MODE	Mode of image read (GRAY or RGB)	str	"RGB"
NN_LEARNING_RATE	The percentage of learning rate	float	1e-4
NN_TRANSFORM_RESIZE	The size of the image transformation	int	224
NN_TRAIN_BATCH_SIZE	The batch size used for training	int	128
NN_TEST_BATCH_SIZE	The batch size used for testing	int	50
NN_VAL_BATCH_SIZE	The batch size used for validation	int	15
NN_NUM_EPOCHS	The number of epochs during training	int	25

Model and Plot

Name	Description	Type	Default
MODEL	Specify which model you want to use ["pretrained" or "custom"]. If missing the script will iterate over both models (first custom, then pretrained)	str	"both"
SAVE_AS_YAML	Saves hyperparameters and results in yaml file	bool	True

Test Resources

The test resources used are images fetched directly from the Stanford40 public dataset.