logo_test/test_results/test_results_analysis.md

Logo Detection Test Results Analysis

This document provides analysis of logo detection test results across different matching methods and configurations.

---

Test Run: CLIP Defaults with All Matching Methods

Date: 2025-12-31 Embedding Model: openai/clip-vit-large-patch14 (default)

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	~2,350
Similarity threshold	0.70
DETR threshold	0.50
Margin	0.05
Min matching refs	3
Random seed	42

Results Summary

Method	TP	FP	FN	Precision	Recall	F1
Simple	751	58,221	9	1.3%	203.5%*	2.5%
Margin	60	26	310	69.8%	16.3%	26.4%
Multi-ref (mean)	233	217	170	51.8%	63.1%	56.9%
Multi-ref (max)	278	259	136	51.8%	75.3%	61.4%

*Recall >100% indicates multiple true positive detections per expected logo (multiple detected regions matching the same logo).

Analysis by Method

Simple Matching

The simple method returns ALL logos above the similarity threshold without any rejection logic. This serves as a baseline to understand the raw discriminative power of CLIP embeddings.

Observations:

• 58,221 false positives vs 751 true positives (~78:1 ratio)
• At threshold 0.70, CLIP embeddings are not discriminative enough to distinguish between different logos
• The extremely high false positive count indicates that unrelated logo regions frequently produce similarity scores above 0.70
• This method is unsuitable for production use but valuable for understanding the embedding space

Margin-Based Matching

The margin method requires the best match to exceed the second-best by a minimum margin (0.05), rejecting ambiguous matches.

Observations:

• Highest precision (69.8%) but very low recall (16.3%)
• Only 60 true positives out of 369 expected
• The margin requirement is too strict when using multiple references per logo
• With 10 refs per logo, references from the SAME logo compete with each other
  • Example: If Logo A has refs scoring 0.85 and 0.84, the margin is only 0.01, causing rejection
• This explains why margin matching produces fewer matches than multi-ref methods

Multi-Ref Matching (Mean Similarity)

Uses the average similarity across all reference images for each logo.

Observations:

• Balanced precision (51.8%) and recall (63.1%)
• F1 score of 56.9%
• False positive ratio approximately 1:1 with true positives (217 FP vs 233 TP)
• Mean aggregation penalizes logos where some references don't match well
• More conservative than max aggregation

Multi-Ref Matching (Max Similarity)

Uses the highest similarity score from any single reference image.

Observations:

• Best F1 score (61.4%) and recall (75.3%)
• Same precision as mean method (51.8%)
• 278 true positives vs 259 false positives (still approximately 1:1)
• Max aggregation is more lenient, improving recall at no precision cost
• Better suited when reference images capture different logo variants

Key Findings

1. CLIP Embedding Similarity Distribution

The simple matching results reveal a fundamental issue: at threshold 0.70, the CLIP embedding space does not provide sufficient separation between different logos. The 78:1 false positive to true positive ratio indicates that:

• Many unrelated images produce high cosine similarity scores
• The threshold would need to be significantly higher (0.85+) to reduce false positives
• Even then, recall would likely suffer

2. Margin Method Limitation with Multiple References

The margin-based matching method was designed assuming one reference per logo. When using multiple references (10 per logo in this test), references from the same logo compete against each other in the margin calculation. This causes legitimate matches to be rejected when two references from the same logo have similar scores.

3. False Positive Rate Remains High

Even the best-performing method (multi-ref max) produces nearly as many false positives as true positives:

• 278 correct matches
• 259 incorrect matches
• This 1:1 ratio is problematic for production use cases

4. Trade-off Between Precision and Recall

Goal	Best Method	Trade-off
Maximize precision	Margin	Very low recall (16.3%)
Maximize recall	Multi-ref (max)	Lower precision (51.8%)
Balance both	Multi-ref (max)	Best F1 but still ~50% precision

Deficiencies of This Approach

CLIP Model Limitations

1. General-Purpose Training: CLIP was trained on text-image pairs for general visual understanding, not for fine-grained logo discrimination. Logo matching requires distinguishing between visually similar brand marks, which CLIP's training objective doesn't optimize for.

2. Embedding Space Density: The cosine similarity scores cluster in a narrow range (0.6-0.9 for most images), making threshold-based discrimination difficult. Small differences in embedding similarity don't reliably indicate visual differences.

3. Scale and Context Sensitivity: CLIP embeddings are affected by the context around detected regions. A logo on a busy background may produce different embeddings than the same logo on a clean background.

4. No Logo-Specific Features: CLIP doesn't learn features specific to logo recognition such as:

   • Typography and font shapes
   • Brand-specific color combinations
   • Geometric patterns and symmetry
   • Edge and contour characteristics

Detection Pipeline Issues

1. DETR Detection Quality: The pipeline assumes DETR correctly identifies logo regions. Detection errors (missed logos, partial detections, non-logo regions) propagate to the matching stage.

2. Cropping Artifacts: Detected regions are cropped and resized before embedding extraction. This may introduce artifacts that affect embedding quality.

3. Threshold Sensitivity: The entire system is highly sensitive to the similarity threshold parameter. A 0.05 change in threshold can dramatically alter precision/recall balance.

---

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.

---

Summary and Recommendations

This section synthesizes findings from all test runs to provide actionable recommendations for logo detection configuration and future improvements.

Best Configuration

Based on all tests conducted, the optimal configuration is:

Parameter	Recommended Value	Rationale
Embedding Model	openai/clip-vit-large-patch14	2x better F1 than DINOv2 alternatives
Matching Method	multi-ref with max similarity	Best F1 (59.9%) and recall (77.0%)
Similarity Threshold	0.70	Lower thresholds outperform higher ones
Margin	0.05	Minimal impact; keep low to avoid rejecting valid matches
Min Matching Refs	3	Provides better discrimination than threshold alone
Refs Per Logo	10	More references improve robustness
DETR Threshold	0.50	Standard detection confidence

Performance Expectations

With the recommended configuration:

Metric	Expected Value	Interpretation
Precision	~49%	About half of detections are correct
Recall	~77%	Finds most logos present in images
F1 Score	~60%	Moderate overall accuracy
FP:TP Ratio	~1:1	Approximately equal true and false positives

Important: These results indicate the system is suitable for applications that can tolerate a high false positive rate, such as:

• Initial screening with human review
• Flagging content for further analysis
• Low-stakes logo presence detection

The system is not suitable for high-precision applications without additional filtering or verification steps.

Key Insights from Testing

What Works

1. Multi-ref matching with max aggregation consistently outperforms other methods
2. Multiple references per logo (10) provides robustness against logo variations
3. min_matching_refs=3 is more effective at discrimination than threshold tuning
4. CLIP embeddings significantly outperform self-supervised alternatives (DINOv2)

What Doesn't Work

1. Raising similarity threshold paradoxically increases false positives in the 0.70-0.85 range
2. Margin-only matching fails with multiple references (same-logo refs compete)
3. DINOv2 models produce 2-3x worse results than CLIP
4. Simple threshold-based matching produces unacceptable 78:1 FP:TP ratio

Limitations

1. ~50% precision ceiling: Even the best configuration produces nearly as many false positives as true positives
2. No clean threshold separation: CLIP's embedding space doesn't provide clear decision boundaries for logos
3. General-purpose models: Neither CLIP nor DINOv2 are optimized for fine-grained logo discrimination
4. Pipeline dependencies: Results depend heavily on DETR detection quality

Recommendations for Future Improvements

Short-Term Improvements

Improvement	Expected Impact	Effort
Post-processing filters	Reduce FP by 20-30%	Low
Add color histogram matching	Filter matches with wrong colors
Add aspect ratio validation	Reject shape mismatches
Add text detection	Filter if expected text is missing
Reference curation	Improve TP by 10-20%	Low
Remove low-quality references	Reduce noise in ref embeddings
Ensure diverse logo variants	Improve coverage
Ensemble scoring	Improve F1 by 10-15%	Medium
Combine CLIP + color features	Multi-signal confidence
Weighted voting across refs	More robust aggregation

Medium-Term Improvements

Improvement	Expected Impact	Effort
Fine-tune CLIP on logos	Improve F1 by 20-40%	Medium
Contrastive training on logo pairs	Better embedding separation
Use LogoDet-3K for training data	Domain-specific features
Alternative detection models	Improve detection quality	Medium
Test YOLOv8 for logo detection	Faster, potentially more accurate
Train custom detector on logo data	Better region proposals
Learned similarity metric	Improve precision by 30-50%	Medium
Train siamese network for logo matching	Replace cosine similarity
Learn logo-specific distance function	Better discrimination

Long-Term Improvements

Improvement	Expected Impact	Effort
End-to-end logo recognition model	F1 > 85%	High
Single model for detection + recognition	Eliminate pipeline errors
Train on large-scale logo dataset	Comprehensive coverage
Logo-specific foundation model	F1 > 90%	High
Pre-train on millions of logo images	Domain expertise
Fine-tune for specific brand sets	Production-ready accuracy

Decision Framework

Use this framework to choose between precision and recall:

Use Case	Priority	Recommended Adjustments
Content moderation	High recall	Use defaults; accept FPs for human review
Brand monitoring	Balanced	Use defaults; filter obvious FPs
Automated licensing	High precision	Use threshold=0.90; accept low recall
Search/discovery	High recall	Lower threshold to 0.65; more refs

Conclusion

The current DETR + CLIP pipeline with multi-ref matching achieves moderate accuracy (~60% F1) that is suitable for screening applications but falls short of production requirements for automated decision-making. The fundamental limitation is that general-purpose vision models lack the fine-grained discrimination needed for logo recognition.

To achieve production-quality accuracy (>85% F1), the system requires:

1. A logo-specific embedding model (fine-tuned or trained from scratch)
2. Additional visual features beyond CLIP embeddings
3. Potentially an end-to-end architecture designed for logo recognition

The test framework established here provides the foundation for evaluating these future improvements systematically.

---

Test Run: [Next Test Name]

Results pending...

---