Player not loading? Watch on YouTube
This tutorial, uploaded on August 24, 2023, walks through RTMDet training with MMDetection in a Google Colab notebook. It uses the large RTMDet variant, RTMDet-L; the supplied material gives no package version numbers. The installation steps describe that dated environment.
The speaker checks GPU access with nvidia-smi, then installs PyTorch and the OpenMMLab dependencies. For a local setup, he directs viewers to choose a PyTorch installation command that matches their operating system, hardware and Python version. The demonstration itself runs in Colab rather than on the viewer's own hardware.
A COCO-pretrained model provides the first inference example. The tutorial explains the roles of model weights and configuration files, then uses Supervision to annotate predictions. When the initial output contains hundreds of boxes, the speaker applies a 0.3 confidence threshold and non-max suppression.
Training uses a domino dataset downloaded from Roboflow Universe. The speaker warns that his MMDetection workflow requires category and image IDs to start at 1, and selects Roboflow's COCO MMDetection export. He edits the configuration with dataset paths, target classes, training parameters and TensorBoard logging.
The reported evaluation reaches roughly 0.7 mAP overall, with one class near 0.4. The speaker suggests checking that class's data and cautions that production images should resemble the training conditions, including lighting and camera angle.
