Training Object Detection Ai with Label Studio and MMDetection

Some time ago I trained [object detector for finding uncapped concrete reinforcement bars] (https://www.glukhov.org/post/2019/object-detection-with-tensorflow/).

Now published a doc with stramlined workflow and detailed guide for setting up an object detection model using Label Studio and Mask RCNN with MMDetection, including the configuration, training, and inference processes.

Here’s a summary of the key steps:

1. Data Preparation:

   • Use Label Studio to create annotations for your dataset.
   • Configure Cloud Storage on Local Files sync
   • Ensure the COCO format is chosen, as it is compatible with MMDetection.

2. Configuring Mask RCNN Model:

   • Customize the base config from mask-rcnn_r50-caffe_fpn_ms-poly-1x_coco.py to adjust for your specific classes and dataset.
   • Modify model parameters like number of classes in head, data root paths, metainfo details, dataloader settings, etc.

3. Training:

   • Execute the training script using MMDetection tool by specifying the config file path, weights directory, and other necessary parameters.

4. Inference:

   • Use DetInferencer from MMDetection for inference on single or multiple images.
   • Specify input files or directories, output directories, and whether to save predictions.

5. Additional Resources:

   • Visit Label Studio site for instructions on setting up this annotation tool: https://labelstud.io/
   • Refer to the official documentation of MMDetection for more details: https://mmdetection.readthedocs.io/en/latest/get_started.html
   • Utilize resources like Bash Cheat Sheet for system operations.

The provided information covers all aspects from preparing data to post-training inference, making it a comprehensive guide for setting up and running Mask RCNN models with MMDetection. Ensure you have correctly set your environment paths and configurations before proceeding with each step.