Deep MAR Lite

Contents

This project is an updated version of Deep-MAR.

The version of PyTorx has been updated to a more recent one.

The net is able to work both in GPU and CPU.

Requirements

The required packages have been version-pinned in the requirements.txt.

torch==1.2 numpy==1.16.4 Pillow pandas

Dockerfile

This version includes two Dockerfiles:

• dev-nogpu.dockerfile
• dev-gpu.dockerfile: CUDA 10 // CUDNN 7

For testing the GPU container, after building it, the user could use the following command:

docker run --runtime=nvidia -it --volume /home/ec2-user/deepMAR:/opt/project deepMAR-dev:gpu /bin/bash

Or a docker-compose file could be used as well.

Weights

Default weights are provided by the author at this issue.

How to run it

The script main.py has the parameters that the user can modify. Unfortunately, in the present version, parameters have to be hard coded in the script.

A future implementation will include a better way of changing the parameters.

The net ingests cropped images, detections, bounding boxes with a requested format. The format is shown in the 1^st point.

1.- Load a pandas dataframe:

# Load pandas dataframe
df = pd.read_csv("./dataset/input_dataset.csv")

One of the columns has to be a string of a dict with the following format

img_dict = {
    "width": image_width,
    "height": image_height,
    "colors": colors,
    "image": zlibed.hex()
}

Where the image "zlibed", in bytes format, has to be converted to hexadecimal format.

2.- Uncompress the cropped image:

# Uncompress cropped image
df["uncompressed_feature_vector"] = df.apply(lambda x: uncompress_string_image(
    compresed_cropped_image=x["feature_vector"]),
    axis=1)

3.- Declare the encoder object, with its correspondent variables and constants.

# Declare an encoder object
# Declare model variables and constants
model_weight_path = "./weights/segmentation_weights.pth"
RESIZE_DIMENSIONS = (224, 224)  # By default
MEAN = [0.485, 0.456, 0.406]
STD = [0.229, 0.224, 0.225]

# Declare an object of the class AttributeExtractor (Model initialization) 
attribute_extractor = AttributeExtractor(
    weight_path=model_weight_path,
    resize_dimensions=RESIZE_DIMENSIONS,
    mean=MEAN,
    std=STD
)

4.- Register the new column as a new dataframe feature:

# Add the new column
df["feature_vector"] = attribute_extractor.get_features(list(df["uncompressed_feature_vector"]))
# Clean the dataframe
df.drop("uncompressed_feature_vector", axis=1, inplace=True)

5.- Save the data as a .csv:

# Write the dataframe to a .csv
df.to_csv("./output_files/output_dataset.csv",
index=False,
quoting=csv.QUOTE_NONNUMERIC
)

Performance & AMI

The attached dockerfiles have been tested on an Amazon Linux AMI 2.0.20190614 x86_64 ECS HVM GP2.

CPU (2,2 GHz Intel Core i7):	GPU (Tesla K80)
10 Hz	54 Hz

Citation

@inproceedings{li2015deepmar,
    author = {Dangwei Li and Xiaotang Chen and Kaiqi Huang},
    title = {Multi-attribute Learning for Pedestrian Attribute Recognition in Surveillance Scenarios},
    booktitle = {ACPR},
    pages={111--115},
    year = {2015}
}