Dataset preparation
Converted VisDrone annotations for training and COCO-style scoring: 6,471 training images, 548 validation images, and 1,610 final test images.
MACHINE LEARNING PROJECT · 2026
Aerial object detection.
Fine-tuning YOLO11n on VisDrone and measuring the improvement over a COCO-pretrained baseline.
2.21% → 15.09%Shared-class mAP@0.5 on 1,610 held-out images.
Python · PyTorch · YOLO11n · pycocotools
Explore the comparison ↓
EVALUATION
Final results.
Both frozen models evaluated once on the same 1,610 test-dev images. Six shared categories, identical inference settings, and explicit COCO-style scoring.
Baseline vs. fine-tuned
| Metric | Baseline | Fine-tuned | Δ pp |
|---|---|---|---|
| mAP @ 0.5 | 2.21% | 15.09% | +12.88 |
| AP | 1.29% | 7.57% | +6.28 |
| Small-object AP | 0.21% | 1.76% | +1.54 |
| Medium-object AP | 2.06% | 11.50% | +9.43 |
| Large-object AP | 10.54% | 24.56% | +14.02 |
AP is mAP@0.5:0.95. Object sizes use original-resolution COCO area ranges. pp = percentage points.
Shared-class performance
| Category | Baseline AP | Fine-tuned AP |
|---|---|---|
| Person | 1.35% | 2.22% |
| Bicycle | 0.00% | 0.05% |
| Car | 3.75% | 21.47% |
| Truck | 1.34% | 5.27% |
| Bus | 1.02% | 14.66% |
| Motorcycle | 0.29% | 1.78% |
What improved—and what remains difficult
All six shared categories improved in AP. Cars and buses show the largest gains, while bicycles and small objects remain difficult. Better than the baseline does not mean reliable enough for deployment.
Explore all 10 VisDrone classes +
This separate task includes van, tricycle, and awning-tricycle. Its scores are not directly comparable to the six-class baseline table.
| Category | AP@0.5:0.95 | AP@0.5 |
|---|---|---|
| Pedestrian | 2.22% | 6.25% |
| People | 0.55% | 1.61% |
| Bicycle | 0.05% | 0.26% |
| Car | 20.11% | 39.66% |
| Van | 3.06% | 4.63% |
| Truck | 5.12% | 9.15% |
| Tricycle | 1.41% | 3.22% |
| Awning-Tricycle | 0.87% | 1.46% |
| Bus | 14.45% | 25.54% |
| Motor | 1.72% | 5.24% |
Full ten-class results
mAP@0.5:0.95 4.96%
mAP@0.5 9.70%
Small-object AP 1.09%
Medium-object AP 7.54%
Large-object AP 18.12%
Project COCO-style metrics, not official VisDrone leaderboard scores. AP summarizes performance across confidence thresholds; it is not the percentage of objects correctly detected. One seed; no statistical-significance claim. Validation results use a different image set and are not mixed into this final comparison.
VALIDATION ERROR ANALYSIS
Compare the predictions.
The same images, shown three ways. Start with the original frames, then explore selected gains and failures.
Six shared classes · confidence ≥ 0.25 for both models · matching IoU ≥ 0.5
Loading comparison data…Lime: ground truth · green: matched · orange: false alarm · gray: ignored
How the cases were selected +
Three original site frames, fixed before inspecting model predictions; then three largest increases in matched objects, three largest decreases, and three largest remaining miss counts. Each additional frame comes from a different sequence prefix; ties use image ID. These selected examples are not representative averages.
Counts use category-aware COCO matching with the same crowd ignore policy as evaluation. The 0.25 confidence cutoff is a fixed display choice, not an optimized operating point. Published AP uses predictions down to 0.001 and remains unchanged. Dashed gray boxes include ignored and excluded-class regions.
Training curves & confusion analysis +
Training losses over 50 epochs. Confusion charts use localization-first matching (IoU ≥ 0.5), then compare classes; their matching differs from the category-aware counts above. Counts are raw; background represents misses or false alarms.
IMPLEMENTATION
What I built.
An end-to-end pipeline for dataset conversion, recoverable training, and comparable evaluation.
Class mapping
Person, bicycle, car, truck, bus, motorcycle. Both models map into the same six-class label space. Full ten-class results are reported separately.
Evaluation
pycocotools scoring at 320px, confidence 0.001, NMS IoU 0.7, and detection caps of 3,000. Category-aware crowd annotations handle ignored regions.
Training & recovery
Seeded training, recorded configurations, environment locks, and Drive checkpoints. Optimizer recovery verified. 25 automated tests passed. Source and checkpoint hashes preserved.
Read the evaluation caveats +
COCO and VisDrone have different taxonomies. Pedestrian and people are merged into person after native class-aware NMS; no second NMS is applied. Van, tricycle, and awning-tricycle are non-scored regions in the shared comparison. Ignored areas are represented as crowd boxes in every evaluated category, a project approximation of VisDrone’s ignore rules. One training seed does not establish statistical significance.
Download evaluation settings ↓Next steps.
Next experiments: higher input resolution or tiled inference for small objects, validation-only threshold calibration, and more training seeds. These are proposed experiments, not measured improvements. Real-time inference and deployment remain unevaluated.