fast_paddle_detection 0.0.3

High-performance offline object detection plugin using PP-PicoDet + NCNN with native camera, anti-spoof, and GPU acceleration.

Changelog

0.0.3

New Features

• 🏷️ Dynamic Labels from File: Load class names from labels.txt
  • loadLabelsFromAsset('labels.txt') — from Android assets
  • loadLabelsFromFile('/path/to/labels.txt') — pick from device storage
  • setLabels(['class0', 'class1', ...]) — set programmatically
  • No hardcoded label names — all driven by external file
  • Pick labels button in settings UI

Bug Fixes

• Fixed num_class always showing 2 for multi-class models (now uses dummy inference at load time)
• Fixed scores/boxes tensor transpose detection (handles both [w=nc, h=anchors] and [w=anchors, h=nc])
• Fixed labels showing COCO names ("person", "bicycle") for custom finetune models
• Fixed picked labels being overridden by asset labels on model reload

Improvements

• Labels auto-loaded from labels.txt in assets at startup
• Custom labels path persisted across model reloads
• Reset labels when switching back to bundled model
• C++ draw_detections now fully dynamic (uses custom_labels vector, falls back to COCO for 80-class, generic "C0" otherwise)

0.0.2

New Features

• 🔀 Multi-Model Support: Auto-detects model format at load time
  • Format A: Models with baked-in post-processing (finetune/custom, 2 inputs)
  • Format C: Models without post-processing (COCO pretrained, 1 input, recommended)
• 🧠 FPN Decode: Full DFL softmax + expectation decode for no-postprocess models
• 🎛️ Runtime Model Loading: Pick .param and .bin separately from device storage
• ⚡ GPU Toggle: Switch CPU/GPU (Vulkan) from settings with auto-detection
• 🛡️ Anti-Spoof: FFT moiré pattern detection to reject screen/monitor images
• 📱 Orientation Aware: Camera preview rotates with device physical orientation
• 📷 Capture Guard: Photo capture only enabled when objects are detected
• ▶️ Start/Stop Camera: Manual control over detection pipeline

Improvements

• Optimized pipeline: reduced from 4 bitmap copies to 1 per frame
• On-demand capture allocation (no per-frame overhead)
• Separate .param and .bin file picker buttons
• GPU availability check before enabling toggle
• Resize → Interp layer replacement documented

Bug Fixes

• Fixed bbox position mismatch in camera (draw in native C++)
• Fixed threshold not applying to all classes equally
• Fixed capture crash (recycled bitmap race condition)
• Fixed camera black screen (runtime permission)
• Fixed landscape detection (OrientationEventListener + targetRotation)
• Fixed preview flip when device rotated
• Fixed GPU switch not responsive in settings sheet
• Fixed model load failure due to mismatched .bin filename
• Fixed Resize layer not supported (replaced with Interp in .param)
• Fixed num_class showing 1600 instead of 80 (min dimension heuristic)

Model Support

• PP-PicoDet S/M/L (320×320)
• COCO 80-class pretrained models (no postprocess)
• Custom finetune models (any number of classes, with postprocess)
• ONNX → NCNN via PNNX conversion documented

0.0.1

• Initial project setup
• Basic NCNN inference pipeline
• JNI bridge + Kotlin plugin
• Dart API + method channel