Darknet CrowdHuman

As part of my Final Year Project, I have trained the Yolo object detector on the Crowdhuman dataset. The goal was to be able to achieve fast detections on people and faces in crowds.

NOTE:

I've realized about a year later that this repo now comes up as one of the top links on a google search for "Crowdhuman", which is nice to see. However, I keep getting emails asking for the weights I trained.

I no longer have access to the weights, they were trained and saved on the Imperial College EEE network drives. I am no longer a student at Imperial College (having graduated thanks to this project), and as such, no longer have access to the network.

TLDR; The weights are lost in the depths of the EEE network oceans.

I've left this repository up as a GUIDE to train the network. Its not too hard, and didnt take too long, about 24 hours on a 4GB GTX 1050Ti

Darknet

Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.

For more information see the Darknet project website.

For questions or issues please use the Google Group.

CrowdHuman Dataset

From the CrowdHuman website:

CrowdHuman is a benchmark dataset to better evaluate detectors in crowd scenarios. The CrowdHuman dataset is large, rich-annotated and contains high diversity. CrowdHuman contains 15000, 4370 and 5000 images for training, validation, and testing, respectively. There are a total of 470K human instances from train and validation subsets and 23 persons per image, with various kinds of occlusions in the dataset. Each human instance is annotated with a head bounding-box, human visible-region bounding-box and human full-body bounding-box. We hope our dataset will serve as a solid baseline and help promote future research in human detection tasks.

Train/Validation

The CrowdHuman dataset can be downloaded from the here.

The training set is divided into 3 files, and are between 2-3GB zipped.

• CrowdHuman_train01.zip
• CrowdHuman_train02.zip
• CrowdHuman_train03.zip

A validation set is also provided in CrowdHuman_val.zip

Both the training and validation sets come with annotations.

• annotation_train.odgt
• annotation_val.odgt

Annotation Format

The annotations come in the .odtg format. Each line in the files is a JSON containing the annotations found in the referred image.

Test

The test set is provided CrowdHuman_test.zip. As far as I can tell, there are no labels for the test set.

Training Darknet

Training on Darknet is never fun. There are about a million different tutorials on how to setup the files, what to do, what files to change. Many tutorials are out of date. I'm about to contribute to this mess.

Note: This tutorial is not meant to be general 'how to'. It is how I managed to train YOLO.

I trained the yolov3-tiny with 2 classes on:

• Ubuntu 16.04 with Standard Darknet Setup (GPU, OPENCV, CUDNN)
• GTX 1050 Ti (4GB RAM)

Setting up files

1. Download the 3 training zip files CrowdHuman_train0*.zip from the CrowdHuman website.
2. Extract all the files into the folder provided /darknet/crowdhuman_train. You should have about 15000 files in the folder.
3. Download the validation zip file CrowdHuman_val.zip
4. Extract the validation set into /darknet/crowdhuman_val. I think theres about 4370 files in there.
5. Download the both annotation_train.odgt and annotation_val.odgt and place them in the main /darknet/ folder.

Crowdhuman to Darknet Format

CrowdHuman provides its annotations in the .odtg JSON format. Darknet does not like this. Darknet expects its annotations as such:

• Each image has a corresponding textfile containing the annotations.

  • Example: dog.jpg would have annotations in dog.txt, in the same folder.

• Annotations in Darknet look like this:

  • <object-class> <x> <y> <width> <height>
  • Each line in the textfile is of that format, and each line represents an object.
  • x, y is the centre of the bounding box.
  • width/height is from the centre of the box.
  • All these values need to be scaled with respect to the size of the image.
    • x = xCentre / imgWidth
    • y = yCentre / imgHeight
    • width = widthBoundingBox / imgWidth
    • height = heightBoundingBox / imgHeight
  • All the values should be between 0 and 1.

Generate Annotations

I have written some files which convert and create the textfiles containing the annotations:

• crowdhuman_train_anno.py
• crowdhuman_val_anno.py

Simply run those two files from the terminal python crowdhuman_*_anno.py from this folder, and it will generate all the corresponding textfiles in the /darknet/crowhuman_train and /darknet/crowhuman_val directories.

Generate Image Filepaths

Darknet also needs another textfile which contains the paths to all the training and validation images. I have written some scripts to do this:

• generate_train_txt.py
• generate_val_txt.py

This generates two files train.txt and val.txt in the main /darknet/ directory.

Actually Training Darknet

Just run the command:

./darknet detector train cfg/yolo_crowdhuman.data cfg/yolov3-tiny-crowdhuman.cfg darknet53.conv.74

Where darknet53.conv.74 is the initial weights which one can get from:

wget https://pjreddie.com/media/files/darknet53.conv.74

Training Tips & Tricks

• No GUI to save GPU Memory
  • Might run out of memory on your GPU, so a good hack is to just run the training without any GUI. I used the virtual terminals tty1. (Or pressing CTRL + ALT + F1)
  • I killed the GUI by running sudo service lightdm stop. This left me with just a terminal and I trained the network there.

Results

View Results:

 ./darknet detector test cfg/yolo_crowdhuman.data  cfg/yolov3-tiny-crowdhuman.cfg backup/yolov3-tiny-crowdhuman_30000.weights Image