Object Detection

YOLO and other single-stage detectors (YOLOv8, YOLOv11) are the default choice for real-time applications because they run one pass over the image and hit 30-60+ fps on modest GPUs. Two-stage detectors like Faster R-CNN or Detectron2's architectures are more accurate on small or overlapping objects but slower, so we reach for them when accuracy matters more than frame rate, such as offline batch analysis rather than a live video feed.

Yes. We export trained models to ONNX or TensorRT and run inference on edge hardware like NVIDIA Jetson boards, Coral Edge TPUs, or modern smartphone NPUs. This adds a step (quantization and hardware-specific optimization) and typically costs some accuracy versus the full-size cloud model, but removes network latency and keeps footage from leaving the device.

Starting from a COCO-pretrained backbone, a few hundred to low thousands of labeled images per object class is a reasonable starting point for transfer learning, though it depends heavily on how visually distinct your classes are and how much variation (lighting, angle, occlusion) they need to handle. We usually train a baseline on an initial batch, measure where it fails, and label more data targeted at those failure modes.

Larger models with higher input resolution generally detect more accurately, especially on small objects, but cost more compute per frame. We manage this through model size selection (e.g. YOLOv8n vs. YOLOv8x), input resolution, and quantization (FP16 or INT8), and tune the confidence threshold to balance false positives against missed detections for your specific use case.