The goal of this assignment is train both models on custom annotated dataset.
-
Take photos of your environment of two or more objects. (at least 100 instances between all objects)
-
Annotate them on roboflow.
-
Train a Faster RCNN model using detectron2
-
Train Yolov4/5/6/7/8 (only one of them of choice) the smallest size
-
Evaluate both models based on mAP and speed and size.
Follow this link: link
I desided to detect spoon and fox. So here are example images:
This is very easy to annotate obhject detection dataset. Here is a screenshot from the roboflow ui:
Based on official documentation of detectron2 and roboflow I was able to train Faster RCNN with detectron2.
Sample predictions:
With the usage of yolo client it was super smooth to train yolov8s on custom dataset:
Sample predictions:
- Mean Average Precision
- Faster RCNN: 72%
- Yolov8: 84.3%
- Speed:
- Yolo is a lot faster and such speed gives a lot of betefits despite its size
- Yolo training for 24 epochs done in 4 minutes, but Faster RCNN in 55 minutes
- Size:
- Faster RCNN model size: 796.5 Mb
- Yolov8 model size: 21.46 Mb