# DLoc Phase 2 Results Summary

## Experiment Setup
- **Dataset**: WILD Jacobs Hall (Fig 10b), 4 APs, 161x361 heatmaps at 0.1m/pixel
- **Training data**: Session 1 (Jul28) + Session 2 train split (Jul28_2) — 17,574 samples
- **Test data**: Session 2 test split — 4,260 samples
- **Hardware**: NVIDIA H100 80GB, PyTorch 2.6.0+cu124
- **Date**: 2026-06-08/09

---

## Step 0: Student Model Training (50 epochs each)

| Model | Params | Median | P90 | P99 |
|-------|--------|--------|-----|-----|
| MobileNet standalone | 2.27M | 0.707m | 1.616m | 3.992m |
| MobileNet distilled | 2.27M | 0.707m | 1.612m | 4.003m |
| TinyCNN standalone | 46.7K | 3.922m | 13.919m | 23.244m |
| TinyCNN distilled | 46.7K | 3.706m | 14.784m | 25.101m |

Distillation loss: L = 0.7 * MSE(student, GT) + 0.3 * MSE(student, teacher_output)

---

## Step 1: Cross-Session Generalization

Models trained on Jul28 sessions only, tested on unseen Aug16 sessions.

| Model | Sess 2 (in-dist) | Sess 3 (furniture) | Sess 4 (diff furn) | Sess 5 (reflector) |
|-------|----------|----------|----------|----------|
| MobileNet standalone | 0.707m | 1.118m | 1.204m | 1.500m |
| MobileNet distilled | 0.707m | 1.118m | 1.204m | 1.442m |
| TinyCNN standalone | 3.922m | 5.964m | 6.327m | 6.605m |
| TinyCNN distilled | 3.706m | 5.852m | 5.903m | 6.888m |
| U-Net cosine standalone | 0.640m | 1.063m | 1.005m | 1.334m |
| **U-Net cosine distilled** | **0.640m** | **1.020m** | **1.005m** | **1.281m** |

U-Net distilled generalizes best across all sessions.

---

## Steps 2 & 3: Robustness Evaluation

### MobileNet Standalone vs Distilled
| Condition | Standalone | Distilled | Improvement |
|-----------|-----------|-----------|-------------|
| Normal | 0.707m | 0.707m | — |
| Random 1-AP dropout | 1.100m | 1.166m | -6% |
| Random 2-AP dropout | 2.402m | 2.516m | -5% |
| Gaussian noise σ=0.1 | 16.368m | 4.769m | **+71%** |
| Gaussian noise σ=0.2 | 16.307m | 5.369m | **+67%** |
| Blur (k=5, σ=1) | 0.849m | 0.854m | — |

Key finding: Distillation improves noise robustness by 3.4x for MobileNet.

### U-Net Cosine (Standalone)
| Condition | Median |
|-----------|--------|
| Normal | 0.640m |
| Random 1-AP dropout | 1.281m |
| Gaussian noise σ=0.1 | 3.000m |
| Gaussian noise σ=0.2 | 5.741m |
| Blur (k=5, σ=1) | 0.671m |

U-Net is the most robust architecture overall.

---

## Step 5: Architectural Improvements

### U-Net + LR Scheduler Comparison (120 epochs)
| Architecture | Scheduler | Mode | Median | P90 | P99 |
|-------------|-----------|------|--------|-----|-----|
| U-Net (2.41M) | Cosine | standalone | **0.640m** | 1.523m | 2.750m |
| U-Net (2.41M) | Step | standalone | **0.640m** | 1.552m | 2.918m |
| U-Net (2.41M) | Cosine | distill | **0.640m** | 1.513m | 3.587m |

Both schedulers converge to same accuracy. Step LR converges faster (epoch 63 vs 110).
U-Net skip connections improve median by 9.5% vs plain MobileNet (0.707m -> 0.640m).

---

## Key Findings

1. **MobileNetV2 matches DLoc baseline** with 7x fewer parameters (2.3M vs 16.5M)
2. **U-Net skip connections** improve accuracy by 9.5% (0.640m vs 0.707m)
3. **Distillation** has minimal impact on accuracy but dramatically improves noise robustness (3.4x for MobileNet)
4. **Cross-session**: U-Net distilled degrades only +0.38-0.64m across environment changes
5. **TinyCNN** (47K params, 350x smaller) achieves 3.7-3.9m — too coarse for practical use
6. **Cosine vs Step LR**: same final accuracy; step converges faster