Still, sticking with YOLO remains the best option considering the 9-hour maximum runtime stipulated for this task.

**1. YOLO (You Only Look Once)**

> Faster — A single-stage detector that predicts bounding boxes and classes in one pass.

> Optimized for real-time detection (typically 30–60 FPS on a good GPU).

> Designed for edge devices (especially YOLOv5 and YOLOv8 with nano or small models).

**2. Faster R-CNN**

❌ Slower — A two-stage detector:

- First proposes regions using a Region Proposal Network (RPN)

- Then classifies and refines bounding boxes

❌ Typically achieves only 5–10 FPS, even on high-end GPUs.

Yeah the train time is much slower for RCNN but I found this paper called "Detecting soybean leaf disease from synthetic image using multi-feature fusion faster R-CNN" where they acheieve decent accuracy with the model. They trained for 10 epochs and got some decent results. When I train for 10 epochs on google colab it takes about 20min per epoch on a T4 instance. The torch implementation of RCNN is also able to export to ONNX which can help with mobile deployment.

For me its YOLO (tried different versions like 11 12 and 9 and different sizes) its working pretty well but am already finding a problem with fitting it into the 9h training frame so I dont know if RCNN will work here.