Performance Metrics
This document provides detailed performance metrics for ESPectre's motion detection algorithms.
Performance Targets
| Scope | Metric | Target | Rationale |
|---|---|---|---|
| MVS / NBVI | Recall | >95% | Minimize missed detections |
| MVS / NBVI | FP Rate | <5% | Avoid false alarms |
| ML | Recall | >95% | All chips use raw std (CV normalization disabled for ML) |
| ML | FP Rate | <5% | Avoid false alarms |
--
Test Configuration
Configuration used for all test results (unified across chips):
| Parameter | Value | Notes |
|---|---|---|
| Window Size | 100 | DETECTOR_DEFAULT_WINDOW_SIZE |
| Calibration | NBVI | Auto-selects 12 non-consecutive subcarriers |
| Hampel Filter | ON | Enabled for both MVS and ML (window=7, threshold=5.0 MAD) |
| Adaptive Threshold | Percentile-based | P95 × 1.1 (DEFAULT_ADAPTIVE_FACTOR) |
| CV Normalization | MVS only | Based on gain_locked metadata (false => apply CV norm for MVS) |
CV normalization is applied per-file for MVS based on whether data was collected with AGC gain lock enabled. ML always uses raw std regardless of gain lock status (the model is trained on raw std).
Training Dataset
micro-espectre/data/dataset_info.json contains multiple baseline/movement captures per chip across different collection sessions and environments. The counts below are aggregated packet totals across all available training captures.
| Chip | Baseline | Movement | Total | Gain Lock |
|---|---|---|---|---|
| ESP32-C3 | 4094 | 4076 | 8170 | Yes |
| ESP32-C5 | 4350 | 4336 | 8686 | Yes |
| ESP32-C6 | 4755 | 4890 | 9645 | Yes |
| ESP32-S3 | 4360 | 4375 | 8735 | Yes |
| ESP32 | 4159 | 4083 | 8242 | No |
| Total | 21718 | 21760 | 43478 | Mixed |
Data location: micro-espectre/data/
Running Tests
source venv/bin/activate
# C++
cd test && pio test -f test_motion_detection -v
# Python (real-data validation)
cd micro-espectre && pytest tests/test_validation_real_data.py::TestPerformanceMetrics -v
# Python (60-second long recordings, prints summary tables)
cd micro-espectre && pytest tests/test_validation_long_recordings.py -v -s
Current Results
Last verified: 2026-05-21 (v2.8.0)
| Chip | Algorithm | Recall | Precision | FP Rate | F1-Score |
|---|---|---|---|---|---|
| ESP32-C3 | MVS Default | 98.5% | 100.0% | 0.0% | 99.3% |
| ESP32-C3 | MVS + NBVI | 96.3% | 100.0% | 0.0% | 98.1% |
| ESP32-C3 | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| ESP32-C5 | MVS Default | 99.4% | 100.0% | 0.0% | 99.7% |
| ESP32-C5 | MVS + NBVI | 99.0% | 100.0% | 0.0% | 99.5% |
| ESP32-C5 | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| ESP32-C6 | MVS Default | 99.7% | 100.0% | 0.0% | 99.9% |
| ESP32-C6 | MVS + NBVI | 99.6% | 100.0% | 0.0% | 99.8% |
| ESP32-C6 | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| ESP32-S3 | MVS Default | 99.7% | 100.0% | 0.0% | 99.9% |
| ESP32-S3 | MVS + NBVI | 99.7% | 100.0% | 0.0% | 99.9% |
| ESP32-S3 | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| ESP32 | MVS Default | 99.4% | 100.0% | 0.0% | 99.7% |
| ESP32 | MVS + NBVI | 99.4% | 100.0% | 0.0% | 99.7% |
| ESP32 | ML | 99.1% | 100.0% | 0.0% | 99.5% |
MVS Default: Uses default subcarriers. MVS + NBVI: Uses NBVI auto-calibration (production case). ML: Neural network with grouped session-level blocked CV for model selection, context-aware MVS-guided weights, and Hampel filtering. CV normalization is always disabled for ML (raw std only).
System Resources
Resource usage benchmarks for ESPectre with full ESPHome stack (WiFi, API, OTA, debug sensors).
Development YAML files (-dev.yaml) include ESPHome debug sensors for runtime monitoring of free heap, max block size, and loop time.
These sensors are available in Home Assistant for continuous monitoring.
Additional performance logs are available at DEBUG level (logger.level: DEBUG):
- [resources] - Free heap at startup and post-calibration
- [perf] - Detection time per packet (logged every ~10 seconds)
Flash Usage
| Chip | Firmware Size | Flash Used | Free App Slot |
|---|---|---|---|
| ESP32-C3 | 1370 KB | 73.8% | 486 KB |
| ESP32-C5 | 1587 KB | 85.5% | 269 KB |
| ESP32-C6 | 1539 KB | 82.9% | 317 KB |
| ESP32-S3 | 1246 KB | 67.1% | 610 KB |
Partition layout uses two app slots (app0/app1, 1.81 MB each) plus a small otadata partition for OTA metadata.
Free App Slot is the remaining space in one app slot after placing the firmware image.
RAM Usage
| Chip | Phase | Free Heap | Notes |
|---|---|---|---|
| ESP32-C3 | Post-setup | 179 KB | After ESPectre init |
| ESP32-C3 | Post-calibration | 83 KB | After NBVI completes |
| ESP32-C5 | Post-setup | 162 KB | After ESPectre init |
| ESP32-C5 | Post-calibration | 71 KB | After NBVI completes |
| ESP32-C6 | Post-setup | 272 KB | After ESPectre init |
| ESP32-C6 | Post-calibration | 180 KB | After NBVI completes |
| ESP32-S3 | Post-setup | 8425 KB | After ESPectre init (includes PSRAM heap) |
| ESP32-S3 | Post-calibration | 8331 KB | After NBVI completes (includes PSRAM heap) |
Detection Timing
Time to process one CSI packet (feature extraction + detection, measured on hardware). At 100 pps, each packet has a 10 ms budget.
| Chip | Algorithm | Detection Time | CPU @ 100 pps |
|---|---|---|---|
| ESP32-C3 | MVS | ~440 µs | ~4.4% |
| ESP32-C3 | ML | ~3400 µs | ~34% |
| ESP32-C5 | MVS | ~220 µs | ~2.2% |
| ESP32-C5 | ML | ~1500 µs | ~15% |
| ESP32-C6 | MVS | ~250 µs | ~2.5% |
| ESP32-C6 | ML | ~1900 µs | ~19% |
| ESP32-S3 | MVS | ~150 µs | ~1.5% |
| ESP32-S3 | ML | ~430 µs | ~4.3% |
The worst-case path is ML on ESP32-C3 (~3.5 ms peak, ~35% CPU), which still leaves substantial budget for WiFi, ESPHome, and Home Assistant communication.
MVS: Extracts a single feature (spatial turbulence) and its moving variance.
ML: Extracts 9 statistical features from sliding window, then runs MLP inference (9 → 32 → 16 → 1 = 816 MACs). The MLP itself is lightweight; most time is spent on feature extraction. For ML architecture details, see ALGORITHMS.md.
60-Second Test Recordings
Continuous recordings (~30s idle + ~30s motion) provide a realistic production-style scenario. These files are not used during training.
Test data: micro-espectre/data/test/
Source of truth: micro-espectre/tests/test_validation_long_recordings.py
Methodology:
- MVS + NBVI: run NBVI calibration on the idle segment, then evaluate the full recording with adaptive threshold and Hampel enabled
- ML: use exported production weights with threshold 5.0 and Hampel enabled
- Both paths skip the first 100 packets of each segment as warmup when scoring packet-level metrics
| Chip | Algorithm | Recall | Precision | FP Rate | F1-Score | FP Count |
|---|---|---|---|---|---|---|
| C3 | MVS + NBVI | 100.0% | 100.0% | 0.0% | 100.0% | 0 |
| C3 | ML | 99.1% | 100.0% | 0.0% | 99.6% | 0 |
| C5 | MVS + NBVI | 100.0% | 91.1% | 7.9% | 95.3% | 253 |
| C5 | ML | 100.0% | 91.1% | 7.9% | 95.3% | 253 |
| C6 | MVS + NBVI | 100.0% | 81.4% | 21.8% | 89.7% | 691 |
| C6 | ML | 89.5% | 96.2% | 3.4% | 92.7% | 108 |
| S3 | MVS + NBVI | 100.0% | 96.9% | 2.8% | 98.4% | 86 |
| S3 | ML | 96.4% | 99.1% | 0.8% | 97.7% | 25 |
Result History (ESP32-C6)
| Date | Version | Dataset | Calibration | Algorithm | Recall | Precision | FP Rate | F1-Score |
|---|---|---|---|---|---|---|---|---|
| 2026-05-21 | v2.8.0 | C6 | - | ML + Hampel | 100.0% | 100.0% | 0.0% | 100.0% |
| 2026-05-21 | v2.8.0 | C6 | NBVI | MVS + Hampel | 99.6% | 100.0% | 0.0% | 99.8% |
| 2026-03-11 | v2.6.1 | C6 | - | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| 2026-03-11 | v2.6.1 | C6 | NBVI | MVS | 99.3% | 100.0% | 0.0% | 99.7% |
| 2026-03-08 | v2.6.0 | C6 | - | ML | 100.0% | 100.0% | 0.0% | 100.0% |
| 2026-03-08 | v2.6.0 | C6 | NBVI | MVS | 99.9% | 98.4% | 2.3% | 99.2% |
| 2026-02-15 | v2.5.0 | C6 | - | ML | 99.9% | 100.0% | 0.0% | 99.9% |
| 2026-02-15 | v2.5.0 | C6 | NBVI | MVS | 99.9% | 99.9% | 0.1% | 99.9% |
| 2026-01-23 | v2.4.0 | C6 | NBVI | MVS | 99.8% | 96.5% | 3.6% | 98.1% |
| 2025-12-27 | v2.3.0 | C6 | NBVI | MVS | 96.4% | 100.0% | 0.0% | 98.2% |
License
GPLv3 - See LICENSE for details.