ExoGlove Object Detection Implementation in Pytorch

This repo implements SSD (Single Shot MultiBox Detector) in PyTorch for object detection, using MobileNet backbones. It also has out-of-box support for retraining on a private ExoGlove dataset and a fruit dataset.

Thanks to @qfgaohao for the upstream implementation from: https://github.com/qfgaohao/pytorch-ssd

Dependencies

1. Python 3.9+
2. OpenCV
3. Pytorch 1.13+
4. Pandas
5. Argparse

Dependencies Installation

1. Install Anaconda
2. Install environment

   conda env create -f environment.yml

3. Activate environment

   conda activate ExoGlove

Run the demo

Run the live demo

CLI Parameters

Parameters	Description	Value Choices
--input-source-path	Absolute Path/Relative Path of the test video

Commend

python run_ssd_example_video.py --input-source-path [PATH_OF_VIDEO]

Test on image

CLI Parameters

Parameters	Description	Value Choices
--input-source-path	Absolute Path/Relative Path of the test image

Commend

python run_ssd_example_image.py --input-source-path [PATH_OF_IMAGE]

Train

CLI Parameters

The ExoGlove-Fruit dataset and the model file can be download from https://sfsu.box.com/s/z2p9usaxxhphpt8s6elkde66oeptqlh6
The dataset need to be place on data folder
The model file need to be place on models folder

Parameters	Description	Value Choices
--net	Network Type(default is mb2-ssd-lite)	mb2-ssd-lite
----pretrained-ssd	Absolute/Relative Path of the pretrained model weight(default is models\mb2-ssd-lite-mp-0_686-pretrained-VOC.pth)
--lr	initial learning rate(default is 0.01)
--momentum	Momentum value for optim(default is 0.9)
--weight-decay	Weight decay for SGD(default is 5e-4)
--gamma	Gamma update for SGD(default is 0.1)
--scheduler	Scheduler for SGD. It can one of multi-step and cosine(default is cosine)	multi-step,cosine
--t-max	T_max value for Cosine Annealing Scheduler(default is 100)
--batch-size	Batch size for training(default is 4)
--num-epochs	the number epochs(default is 200)
--num-workers	Number of workers used in dataloading(default is 2)
--use-cuda	Use CUDA to train model(default is true)
----model-dir	Directory for saving checkpoint models(default is models/my_model)

Commend

python train_ssd.py

Caution

If you are a windows/Mac User, please set the num-workers to 0.

Convert to ONNX Model

CLI Parameters

Parameters	Description	Value Choices
--net	Network Type(default is mb2-ssd-lite)	mb2-ssd-lite
--input	Absolute/Relative Path of the model weight(default is models\model_exoglove_07262022.pth)
--output	Absolute/Relative Path of the output onnx file(default is models\model_exoglove_07262022.onnx)
--labels	Absolute/Relative Path of the labels file(default is models\labels.txt)

commend

python export_onnx.py

Deploy on Jetson Nano

Refer to Locating Objects with DetectNet