Settings

Film Tool

Setup for manual tagging lives here so the Film Tool top bar stays focused on watching and tagging. Use Tagger for day-to-day work; open these when you need vocabulary, export, or appearance options.

Tagging vocabulary Saved games & export Reports Open Film Tool

Film Tool appearance

Saved in this browser only.

Current Hardware

GPU

NVIDIA GeForce RTX 5060 Laptop GPU

GPU Memory

139 / 8151 MB

Packages

OpenCV: Yes
Ultralytics: Yes
Torch: Yes

Local LLMs

None detected

Analysis Behavior

Vision Runtime

Detector Model

yolov8n.pt: Fastest option and the safest default on any hardware.

yolov8s.pt: Good balance of speed and accuracy on GPU, slower on CPU.

yolov8m.pt: Heavier model; best reserved for stronger GPUs or slower offline runs.

yolo11n.pt: Newer lightweight Ultralytics detector with a similar footprint to YOLOv8 Nano.

yolo11s.pt: Good next step if you want to compare a newer small detector against YOLOv8 Small.

yolo11m.pt: Higher quality option for slower offline reruns on stronger GPUs.

custom: Use any supported model name or local .pt path, including your own fine-tuned basketball weights.

Custom Ultralytics Model or Weights

Used only when Detector Model is set to custom. Ollama models are for reasoning, not frame-by-frame detection.

Ball Detector Model

models/ball_detector.pt: Benchmark default for ball detection: class 0 at 0.15 confidence.

yolov8n.pt: Uses COCO class 32 with legacy filters; benchmarked at 0% precision and 0% recall at conf=0.15.

custom: Use a local Ultralytics .pt file for future basketball detector candidates.

Custom Ball Detector Weights

Used only when Ball Detector Model is set to custom.

Ball Class ID

Ball Confidence Threshold

Person Confidence Threshold

Inference Device

Auto-Accept Event Confidence

AI events at or above this confidence are auto-accepted after analysis. Set to 0 to disable.

Event Generator

Legacy generator: Basic event generation from detections.

Expanded heuristic generator: Recommended. Builds on the current detections to emit possession changes, shots, makes, misses, rebounds, assists, steals, turnovers, blocks, and fouls.

Frame Processing Rate

Tracker Max Distance

Tracker Max Frame Gap

Local AI Models

Provider

Model

These models are detected from the local Ollama install and saved for future AI-assisted workflows.

Recommended Ollama Models

Qwen 2.5 3B (Recommended)

qwen2.5:3b — Best lightweight general-purpose local model on this machine.

Llama 3.2 3B (Recommended)

llama3.2:3b — Fast local model for summaries and quick prompts.

Gemma 3 4B (Recommended)

gemma3:4b — Good small-model option if you want a different family than Llama/Qwen.

Qwen 2.5 7B

qwen2.5:7b — Heavier but still reasonable for this hardware if you want stronger quality.

Mistral 7B

mistral:7b — Good mid-sized fallback for analysis and comparisons.

Llama 3.1 8B

llama3.1:8b — Largest practical local model for this setup; already usable but heavier.