Performance Guide
Fovux v1.0.0 focuses on predictable local performance for long-running YOLO workflows.
Metric streaming
The Studio dashboard now prefers metrics.jsonl written directly by the worker. The HTTP stream no longer re-parses the full history every polling tick. Instead:
- a snapshot is sent on connect
- file changes are watched
- only appended rows are streamed after the snapshot
This keeps steady-state CPU usage low for active dashboards.
Dataset inspection
Large datasets benefit from stable mtimes and cached filesystem metadata. When repeatedly inspecting the same dataset:
- keep labels under a consistent root
- avoid rewriting annotation files unnecessarily
- prefer SSD-backed
FOVUX_HOMEfor cache-heavy workflows
Duplicate detection
Perceptual hashing is CPU-bound. For the fastest runs:
- store input images on local disk instead of network storage
- avoid background antivirus scans over the dataset root
- run duplicate scans before opening multiple heavy dashboard sessions
RTSP inference
The RTSP path now uses reconnect backoff and a bounded capture queue. For stable streams:
- keep
CAP_PROP_BUFFERSIZE=1 - prefer streams that report FPS correctly
- write outputs to a fast local disk when recording
Benchmarking
Use benchmark_latency after export or quantization to compare checkpoints in the same environment. Small ONNX and INT8 wins are meaningful only when measured on the target hardware profile.