Object detection with customized architecture

In this tutorial, we demonstrate how to use JustDeepIt for object detection using a user customized neural network architecture.

Dataset

The artificial dataset used for this quick start is stored in GitHub (JustDeepIt/tutorials/OD). The data folder contains images folder, annotations folder, and class_labels.txt file. The images folder contains training images. The annotations folder contains annotation of training images. class_labels.txt is an text file containing the class labels of training images. Users can use git command to download the dataset from GitHub with the following script.

git clone https://github.com/biunit/JustDeepIt

ls JustDeepIt/tutorials/OD
# data run_fasterrcnn.py

ls JustDeepIt/tutorials/OD/data
# annotations     class_label.txt images

ls JustDeepIt/tutorials/OD/data/images
# e1.png  e10.png e2.png  e3.png  e4.png  e5.png  e6.png  e7.png  e8.png  e9.png
../_images/quickstart_od_data.png

Settings

To start JustDeepIt, we open the terminal, as the following commands, change the current directory to JustDeepIt/tutorials/OD, and run justdeepit command.

cd JustDeepIt/tutorials/OD

justdeepit
# INFO:uvicorn.error:Started server process [61]
# INFO:uvicorn.error:Waiting for application startup.
# INFO:uvicorn.error:Application startup complete.
# INFO:uvicorn.error:Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)

Then, we open the web browser and accesss to http://127.0.0.1:8000. At the startup screen, we press “Object Detection” button to start object detection mode.

../_images/app_main.png

Next, at the Preferences screen, we set the architecture to custom, set config to faster_rcnn_x101_64x4d_fpn_mstrain_3x_coco, and the other parameters as shown in the screenshot below.

../_images/tutorials_od_cus_pref.png

The user can find the file to input the config argument from the configs directory of the MMDetection or Detectron2 GitHub repositories. For example, the faster_rcnn of MMDetection repository contains many variants of Faster R-CNN, as follows:

  • faster_rcnn_r50_fpn_mstrain_3x_coco.py

  • faster_rcnn_r101_caffe_fpn_mstrain_3x_coco.py

  • faster_rcnn_x101_32x8d_fpn_mstrain_3x_coco.py

  • …

We can choose one of them to set the config argument of JustDeepIt. Note that when inputing the config argument, the extension .py can be omitted. The screenshot above shows an example of setting Faster R-CNN with ResNeXt backbone (faster_rcnn_x101_32x8d_fpn_mstrain_3x_coco) for object detection.

Then, we press button Load Workspace. Once the Preferences is set, the functions of Training and Inference become available.

From here, the same procedure as in the quick start of object detection can be used for model training and inference. Refer to Training and Inference in the quick start for details.

Training

See Training in the the quick start of object detection for details.

Inference

See Inference in the the quick start of object detection for details.