Refined trash sorting with moving conveyor belts

This is the semester thesis deriving from the ARC project. The whole project is composed of two parts. The first part is to train neural network to detect multiple sorts of trash collected from water. To improve the classifier detector, novel network architecture Swin Transformer has been introduced.

The second part focuses on extracting information from the detected object. Basic structure is realized by python on ROS. Presentation link.

Training Swin Transformer

The training procedure is based on MMDetection developed and maintained by senseTime. This platform integrates plenty of open source algorithms for object detection that could be simply tuned for future adjustment.

Dependencies

Ubuntu 20.04.3 LTS
Nvidia Driver: 470.57.02    
CUDA: 11.4
python:3.7.11 
pytorch: 1.8.1
mmcv :1.3.15
mmdet:2.17.0

Nvidia GeForce GTX 2080

Environment Setup

Prerequisite:
-CUDA 9.2+
-GCC  5+

Step 1.Install prerequisite(skip this if they are already installed)

CUDA reference(select preferable version here)

# install CUDA ,take latest version as example 

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin

sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600

wget https://developer.download.nvidia.com/compute/cuda/11.6.0/local_installers/cuda-repo-ubuntu2004-11-6-local_11.6.0-510.39.01-1_amd64.deb

sudo dpkg -i cuda-repo-ubuntu2004-11-6-local_11.6.0-510.39.01-1_amd64.deb

sudo apt-key add /var/cuda-repo-ubuntu2004-11-6-local/7fa2af80.pub

sudo apt-get updatesudo apt-get -y install cuda

# install GCC
sudo apt update
sudo apt install build-essential
sudo apt-get install manpages-dev

Step 2. Install MMDetection

Official document here (Note: match every version carefully)ss

#1.Create a conda virtual environment and activate it.

conda create -n openmmlab python=3.7 -y
conda activate openmmlab

#2.Install PyTorch and torchvision of suitable version,e.g.

conda install pytorch=1.3.1 cudatoolkit=9.2 torchvision=0.4.2 -c pytorch

#3.Install MMDetection

pip install openmim
mim install mmdet

#or
# install the latest mmcv

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.6.0/index.html

# install mmdetection

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -r requirements/build.txt
pip install -v -e .

Dataset

TACO+TrashNet+Experiment Raw Image

TACO
TrashNet
Experiment

Data Preprocess

remap_classes.py

reads annotaions (.json files) and class remap (.csv file) and outputs new annotation files

python remap_classes.py --class_map maps/map_to_2.csv --ann_dir ./data/annotations.json
merge.py

merge two coco-format annotation files

python merge.py Json1.json Json2.json OUTPU_JSON.json
fix_category.py

fix the category list order to be consistent with trained neural network
split_dataset.py

takes an annotations file (.json) and splits it into training, testing, validation datasets for x many equivalent sets for cross validation.

python split_dataset.py --nr_trials 5 --out_name ann_05 --dataset_dir data
change_subdirectory.py

change the subdirectory name in the annotations file.

python change_subdirectory.py --input new_directory --original annotations.json --new output.json
add_subdirectory.py

add subdirectory into the annotation file.
filter.py

select interested categories from the original annotation file,e.g.

python filter.py --input_json instances_train2017.json --output_json filtered.json --categories person dog cat

Training

official training tutorial link

There is no official support for the cross validation algorithm. If you want to train with 5-fold method, use bash ./train4.sh and change the work-dir in the bash file.

configuration of annotations

model : adjust the roi_head number consistent with the classes you declare

data : specify the train and validation and test data file here

classes: categories for detection

Labeling tools

Note: After labeling, check the correspondence between classes in config and annotationsfile.

Extracting object information

This is the ROS implementation of the object detection system. The whole node structure is as follows.

Once the detector has been launched, the basic information of the object including annotation details and tracking ID will be extracted and saved for future retraining and database statistics.

The example visualization of output is like the following picture.

Prerequisite

Install ROS noetic. Tutorial

Quick Start

conda activate openmmlab

catkin_make

roslaunch swin_detection launch.launch

Note: if there is any package not existed, just install ros packages according to the log under openmmlab environment.

Output

/output_database:cropped images and corresponding .json file

/output_multi:images and annotations_multi.json

/output:single object annotation that is used for debugging

Files Breakdown

/scripts

image_info.py : Define classes for annotation extracting

sort.py: Tracker based on Kalman filter to keep tracking bounding box

image_process_node.py: Core node to deploy the neural network and save multiple object annotation details

object_info_subsribe_node.py: Helping node to extract single object information for verification and debugging

database_save_node.py: Obtain cropped image for every object and save information for database

/launch

launch.launch:launch above nodes and specify the rosbag data path here.

/msg

Object.msg : Customize object message to convey extracted information for sequential nodes

/config

parameters.yaml: load parameters for the detection system

crepuscularlight / ARC_SWIN_DETECTION