Add embedding model comparison analysis (CLIP vs DINOv2)

- Add section explaining how margin works differently in multi-ref vs
  margin-only matching, with examples showing why margin-only fails
  when using multiple references per logo
- Update run_model_comparison.sh to use optimal threshold (0.70) and
  margin (0.05) based on test results
- Add DINOv2 Large model test to comparison script
- Add threshold optimization test analysis to results document

logo_detection_test_methodology.md

@@ -224,6 +224,51 @@ This ensures confident matches and reduces ambiguous classifications.
 - Margin required: 0.05
 - Result: **No match** (0.82 - 0.79 = 0.03 < 0.05)
 
+#### Margin in Multi-Ref vs Margin-Only Matching
+
+The margin parameter applies to both `margin` and `multi-ref` methods, but operates at different levels:
+
+| Method | What Margin Compares |
+|--------|---------------------|
+| `margin` | Best **reference embedding** vs second-best **reference embedding** |
+| `multi-ref` | Best **logo's aggregated score** vs second-best **logo's aggregated score** |
+
+This distinction is critical when using multiple references per logo.
+
+#### The Problem with Margin-Only and Multiple References
+
+In margin-only matching, all individual reference embeddings compete against each other—including references from the **same logo**. This causes legitimate matches to be rejected.
+
+**Example showing the problem:**
+
+Suppose Nike has 3 references and Adidas has 3 references. A detected region produces:
+
+| Reference | Similarity |
+|-----------|------------|
+| Nike_ref1 | 0.92 |
+| Nike_ref2 | 0.91 |
+| Nike_ref3 | 0.85 |
+| Adidas_ref1 | 0.78 |
+| Adidas_ref2 | 0.75 |
+| Adidas_ref3 | 0.72 |
+
+**With margin-only matching (margin=0.05):**
+- Best reference: Nike_ref1 (0.92)
+- Second-best reference: Nike_ref2 (0.91) ← Same logo!
+- Margin check: 0.92 - 0.91 = 0.01 < 0.05 → **Rejected**
+
+The match is rejected even though this is clearly a Nike logo! Nike's own references compete against each other and fail the margin test.
+
+**With multi-ref matching (margin=0.05):**
+- First, aggregate scores per logo:
+  - Nike: max(0.92, 0.91, 0.85) = 0.92
+  - Adidas: max(0.78, 0.75, 0.72) = 0.78
+- Best logo: Nike (0.92)
+- Second-best logo: Adidas (0.78)
+- Margin check: 0.92 - 0.78 = 0.14 >= 0.05 → **Accepted**
+
+This is why margin-only matching produces very low recall when using multiple references per logo—it was designed for single-reference scenarios.
+
 ---
 
 ### 6. Embedding Caching

run_model_comparison.sh

@@ -13,8 +13,8 @@ REFS_PER_LOGO=10
 POSITIVE_SAMPLES=20
 NEGATIVE_SAMPLES=100
 MIN_MATCHING_REFS=3
-THRESHOLD=0.80
+THRESHOLD=0.70
-MARGIN=0.10
+MARGIN=0.05
 SEED=42
 
 # Clear output file and write header
@@ -82,11 +82,29 @@ uv run python "$SCRIPT_DIR/test_logo_detection.py" \
     --clear-cache \
     --output-file "$OUTPUT_FILE"
 
+echo ""
+
+# Test 3: DINOv2 Large
+echo "=== Test 3: DINOv2 Large (facebook/dinov2-large) ==="
+uv run python "$SCRIPT_DIR/test_logo_detection.py" \
+    --num-logos $NUM_LOGOS \
+    --refs-per-logo $REFS_PER_LOGO \
+    --positive-samples $POSITIVE_SAMPLES \
+    --negative-samples $NEGATIVE_SAMPLES \
+    --matching-method multi-ref \
+    --min-matching-refs $MIN_MATCHING_REFS \
+    --use-max-similarity \
+    --threshold $THRESHOLD \
+    --margin $MARGIN \
+    --seed $SEED \
+    --embedding-model "facebook/dinov2-large" \
+    --clear-cache \
+    --output-file "$OUTPUT_FILE"
+
 echo ""
 echo "Results saved to: $OUTPUT_FILE"
 echo ""
 echo "Note: You can also try other models:"
 echo "  - facebook/dinov2-base"
-echo "  - facebook/dinov2-large"
 echo "  - openai/clip-vit-base-patch32"
 echo "  - openai/clip-vit-large-patch14-336"

test_results_analysis.md

@@ -137,6 +137,215 @@ Even the best-performing method (multi-ref max) produces nearly as many false po
 
 ---
 
+## Test Run: Threshold Optimization Tests
+
+**Date**: 2026-01-02
+**Embedding Model**: openai/clip-vit-large-patch14
+**Matching Method**: Multi-ref (max) for all tests
+
+### Test Configuration
+
+| Parameter | Value |
+|-----------|-------|
+| Reference logos | 20 |
+| Refs per logo | 10 |
+| Total reference embeddings | 189 |
+| Positive samples per logo | 20 |
+| Negative samples per logo | 100 |
+| Test images processed | ~2,355 |
+| DETR threshold | 0.50 |
+| Min matching refs | 3 |
+| Random seed | 42 |
+
+### Results Summary
+
+| Test | Threshold | Margin | TP | FP | FN | Precision | Recall | F1 |
+|------|----------:|-------:|---:|---:|---:|----------:|-------:|---:|
+| 1 (baseline) | 0.70 | 0.05 | 265 | 288 | 141 | 47.9% | 71.8% | 57.5% |
+| 2 | 0.80 | 0.05 | 233 | 472 | 165 | 33.0% | 63.1% | 43.4% |
+| 3 | 0.80 | 0.10 | 187 | 375 | 208 | 33.3% | 50.7% | 40.2% |
+| 4 | 0.85 | 0.10 | 160 | 434 | 223 | 26.9% | 43.4% | 33.2% |
+| 5 | 0.85 | 0.15 | 163 | 410 | 220 | 28.4% | 44.2% | 34.6% |
+| 6 | 0.90 | 0.15 | 84 | 69 | 288 | 54.9% | 22.8% | 32.2% |
+
+### Analysis
+
+#### Counter-Intuitive Results
+
+The most striking finding is that **raising the similarity threshold made performance worse** in most cases:
+
+| Threshold Change | Effect on FP:TP Ratio |
+|------------------|----------------------|
+| 0.70 → 0.80 | 1.09:1 → 2.03:1 (worse) |
+| 0.80 → 0.85 | 2.03:1 → 2.71:1 (worse) |
+| 0.85 → 0.90 | 2.71:1 → 0.82:1 (better) |
+
+This is the opposite of expected behavior. Normally, raising the threshold should reduce false positives. Instead, false positives *increased* from 288 at threshold 0.70 to 472 at threshold 0.80.
+
+#### Why Higher Thresholds Failed
+
+The likely explanation relates to how `min_matching_refs` interacts with the threshold:
+
+1. **True positives are penalized more**: Correct matches require 3+ references to exceed the threshold. At higher thresholds, fewer references clear the bar, causing legitimate matches to fail the `min_matching_refs=3` requirement.
+
+2. **False positives survive differently**: False positive detections may have 1-2 references that happen to score very high (above the threshold) due to random visual similarities. Since we use max aggregation, these spurious high scores still produce matches.
+
+3. **The margin becomes less effective**: When most scores are clustered below the threshold, the margin check operates on a smaller pool of candidates, reducing its discriminative power.
+
+#### Threshold 0.90: Different Behavior
+
+At threshold 0.90, behavior finally matches expectations:
+- False positives dropped dramatically (69 vs 288-472 in other tests)
+- But recall collapsed to 22.8%
+- Only 84 true positives out of 369 expected
+
+This suggests that at 0.90, the threshold is finally high enough to filter out most noise, but it's too aggressive and rejects most legitimate matches as well.
+
+#### The Optimal Threshold Problem
+
+| Threshold | Precision | Recall | F1 | Assessment |
+|-----------|-----------|--------|-----|------------|
+| 0.70 | 47.9% | 71.8% | **57.5%** | Best overall F1 |
+| 0.80 | 33.0% | 63.1% | 43.4% | Worse than baseline |
+| 0.85 | 26.9-28.4% | 43-44% | 33-35% | Much worse |
+| 0.90 | 54.9% | 22.8% | 32.2% | Best precision, worst recall |
+
+The lowest threshold tested (0.70) produced the best F1 score. This indicates:
+- CLIP embeddings don't provide clean separation at any threshold
+- The multi-ref matching with min_matching_refs provides better discrimination than threshold alone
+- Raising the threshold hurts true positives more than it helps reject false positives
+
+#### Margin Parameter Impact
+
+Comparing tests with the same threshold but different margins:
+
+| Threshold | Margin 0.05 | Margin 0.10 | Margin 0.15 |
+|-----------|-------------|-------------|-------------|
+| 0.80 | F1: 43.4% | F1: 40.2% | - |
+| 0.85 | - | F1: 33.2% | F1: 34.6% |
+
+Increasing the margin had minimal effect, slightly reducing both true and false positives. The margin parameter is less impactful than the threshold in this configuration.
+
+### Key Findings
+
+1. **The baseline (threshold=0.70, margin=0.05) was optimal**: No threshold/margin combination tested outperformed the defaults for F1 score.
+
+2. **Threshold tuning alone cannot fix CLIP's limitations**: The embedding space doesn't provide clear separation points that can be exploited with threshold adjustments.
+
+3. **min_matching_refs matters more than threshold**: The requirement for multiple matching references provides better discrimination than similarity threshold.
+
+4. **Precision-recall trade-off is extreme**: Achieving 55% precision (at threshold 0.90) requires accepting only 23% recall.
+
+5. **The 0.70-0.85 range is a "dead zone"**: Thresholds in this range produce worse results than either extreme.
+
+### Implications
+
+These results suggest that improving logo detection accuracy requires:
+- A different embedding model with better logo discrimination
+- Logo-specific fine-tuning
+- Alternative matching strategies beyond threshold-based approaches
+- Potentially ensemble methods combining multiple signals
+
+Simply tuning threshold and margin parameters with CLIP is insufficient to achieve acceptable precision/recall balance.
+
+---
+
+## Test Run: Embedding Model Comparison
+
+**Date**: 2026-01-02
+**Matching Method**: Multi-ref (max) for all tests
+
+### Test Configuration
+
+| Parameter | Value |
+|-----------|-------|
+| Reference logos | 20 |
+| Refs per logo | 10 |
+| Total reference embeddings | 189 |
+| Positive samples per logo | 20 |
+| Negative samples per logo | 100 |
+| Test images processed | ~2,355 |
+| Similarity threshold | 0.70 |
+| DETR threshold | 0.50 |
+| Margin | 0.05 |
+| Min matching refs | 3 |
+| Random seed | 42 |
+
+### Results Summary
+
+| Model | TP | FP | FN | Precision | Recall | F1 |
+|-------|---:|---:|---:|----------:|-------:|---:|
+| CLIP ViT-Large | 284 | 295 | 124 | 49.1% | 77.0% | 59.9% |
+| DINOv2 Small | 158 | 546 | 234 | 22.4% | 42.8% | 29.5% |
+| DINOv2 Large | 105 | 221 | 277 | 32.2% | 28.5% | 30.2% |
+
+### Analysis
+
+#### CLIP Significantly Outperforms DINOv2
+
+CLIP ViT-Large achieved approximately **2x the F1 score** of either DINOv2 model:
+
+| Model | F1 Score | vs CLIP |
+|-------|----------|---------|
+| CLIP ViT-Large | 59.9% | baseline |
+| DINOv2 Small | 29.5% | -50.7% |
+| DINOv2 Large | 30.2% | -49.6% |
+
+This is a substantial performance gap that cannot be closed through parameter tuning.
+
+#### DINOv2 Model Comparison
+
+Comparing the two DINOv2 variants:
+
+| Metric | DINOv2 Small | DINOv2 Large | Winner |
+|--------|--------------|--------------|--------|
+| Precision | 22.4% | 32.2% | Large (+44%) |
+| Recall | 42.8% | 28.5% | Small (+50%) |
+| F1 | 29.5% | 30.2% | Large (+2%) |
+| FP:TP Ratio | 3.46:1 | 2.10:1 | Large |
+
+DINOv2 Large shows better precision and fewer false positives, but at the cost of significantly lower recall. The larger model appears more conservative in its matching, rejecting more candidates overall.
+
+#### Why DINOv2 Underperforms
+
+1. **Training Objective Mismatch**: DINOv2 uses self-supervised learning optimized for general visual representation, not for discriminating between similar visual objects. While it excels at semantic understanding, logo matching requires fine-grained visual discrimination.
+
+2. **Embedding Space Characteristics**: DINOv2's embedding space may cluster logos differently than CLIP. The 0.70 threshold that works reasonably for CLIP may be entirely wrong for DINOv2's similarity distribution.
+
+3. **No Text-Image Alignment**: Unlike CLIP, DINOv2 has no concept of semantic labels. CLIP's text-image training may inadvertently help it distinguish between branded content, even if not explicitly trained for logos.
+
+#### False Positive Analysis
+
+| Model | FP:TP Ratio | Assessment |
+|-------|-------------|------------|
+| CLIP ViT-Large | 1.04:1 | Approximately balanced |
+| DINOv2 Small | 3.46:1 | Very high false positives |
+| DINOv2 Large | 2.10:1 | High false positives |
+
+DINOv2 Small produces over 3x as many false positives as true positives, making it unsuitable for this task without significant threshold adjustment.
+
+### Key Findings
+
+1. **CLIP remains the best choice**: Despite its limitations documented in earlier tests, CLIP substantially outperforms DINOv2 for logo matching with the current pipeline and parameters.
+
+2. **Model size doesn't guarantee better results**: DINOv2 Large (304M parameters) performed only marginally better than DINOv2 Small (22M parameters) for F1 score, and actually had worse recall.
+
+3. **Threshold may need per-model tuning**: The 0.70 threshold optimized for CLIP may not be appropriate for DINOv2. The high false positive rates suggest DINOv2 may need a higher threshold.
+
+4. **Self-supervised models not ideal for this task**: The results suggest that self-supervised vision models like DINOv2 are not well-suited for fine-grained logo discrimination without additional fine-tuning.
+
+### Recommendations
+
+1. **Continue using CLIP** for this logo detection pipeline unless a logo-specific model becomes available.
+
+2. **If DINOv2 must be used**, conduct threshold optimization tests specifically for DINOv2's embedding space—the optimal threshold is likely different from CLIP's.
+
+3. **Consider fine-tuning**: Training a model specifically on logo discrimination tasks would likely outperform both general-purpose models.
+
+4. **Explore hybrid approaches**: Combining CLIP's semantic understanding with additional visual features (edges, colors, shapes) might improve discrimination.
+
+---
+
 ## Test Run: [Next Test Name]
 
 *Results pending...*