Uses pipeline script

COCO output format

Correct split ratios

Stratified split

Seed 42

Mosaic augmentation

Mixup augmentation

Cutout parameters

Horizontal flip probability

Rotate limit

Brightness/contrast limits

Hue saturation values

pycocotools verification

YOLOv8 or RT-DETR chosen

Uses vision_model_trainer.py

YOLO CLI training command

mAP@50 threshold

mAP@50:95 threshold

Precision threshold

Recall threshold

Inference time target

CNN vs ViT data reasoning

Detection task flag

Uses inference_optimizer.py

Baseline benchmark step

Benchmark parameters

Dynamic batch flag

Simplify flag

ONNX verification

Cloud path: TensorRT FP16

Intel path: OpenVINO

ONNX intermediary for Intel

INT8 calibration samples

Opset version 17

Instance seg architecture

Real-time instance choice

Semantic seg architecture

SegFormer speed advantage

SAM for zero-shot

SAM prompt types

segment-anything package

Mask R-CNN quality trade-off

mmsegmentation or torchvision

Architecture decision format

Instance vs semantic distinction

ByteTrack or SORT

YOLOv8 detector

Real-time FPS target

Latency P99 target

GPU memory target

Model size target

Persistent track IDs

Video capture loop

Detection then tracking flow

Script is executable

FPS measurement

Edge architecture choice

SAHI or high-res for small objects

CNN over ViT justification

Edge FPS target

Edge mAP@50 target

Edge GPU memory target

Edge model size target

NVIDIA edge optimization path

Copy-paste for small objects

Architecture plan document

P2 FPN level or anchor adjustment