You will get a YOLO + SAHI small-object detection pipeline with measurable mAP gain
Rising Talent
Rising Talent
Project details
I build a YOLO + SAHI small-object detection pipeline that recovers missed tiny targets (drones, debris, traffic, remote sensing) while keeping the run clear and reproducible. I run a parameter sweep (tile size, overlap, NMS/score thresholds) and, in higher tiers, add WBF and fine-tuning to maximize mAP/F1 at a fixed latency. You receive a concise report (PR curves, tables), a clean environment (requirements.txt / env.yml), and a one-command run. Everything is delivered chat-only with fast async feedback. If you share a small data sample and your current run, I’ll return a concrete plan and first improvements quickly.
Machine Learning Tools
Microsoft Excel, MLflow, NumPy, NVIDIA AI Platform, OpenCV, pandas, Python, PyTorch, scikit-learn, SciPy, SQL, TensorFlowWhat's included
| Service Tiers |
Starter
$129
|
Standard
$249
|
Advanced
$399
|
|---|---|---|---|
| Delivery Time | 2 days | 3 days | 5 days |
Number of Revisions | 1 | 2 | 3 |
Number of Model Variations | 0 | 1 | 2 |
Number of Scenarios | 0 | 1 | 2 |
Number of Graphs/Charts | 1 | 2 | 3 |
Model Validation/Testing | |||
Model Documentation | - | ||
Data Source Connectivity | - | - | - |
Source Code | - |
Optional add-ons
You can add these on the next page.
Additional Revision
+$49
Extra mAP tuning iteration
+$79
Export to ONNX/TensorRT
+$79
12h Rush Delivery (custom add-on)
+$99Frequently asked questions
About Dejan
PhD, Senior ML Engineer | YOLO+SAHI CV | LLMs (RAG, QLoRA)
Vranje, Serbia - 5:01 am local time
• Computer Vision: YOLOv5/8 + SAHI + WBF for small-object UAV imagery; +24% mAP@0.50:0.95 with FN↓ at a defined operating point. Real-time PyTorch/OpenCV/CUDA pipelines.
• NLP / LLMs: Led a bilingual GPT-2–class model effort; QLoRA fine-tuning, RAG (Elasticsearch/FAISS), cross-encoder reranking, evaluation (BLEU/ROUGE/BERTScore).
• Ops & Quality: Reproducible repos (Docker, W&B, CI), deterministic runs, clear metric reports. Own GPU server.
• Domains: Legal-tech, safety-critical CV (UAV/remote sensing), time-series forecasting, academia.
• Deliverables: one-command run + configs, README with stability notes, PR curves & tables, short metric brief, post-delivery support window.
• Collaboration: Upwork Messages (chat-only), fixed-price milestones for speed & predictability. Availability: 30+ hrs/week, 24–48h turnaround per revision wave.
Steps for completing your project
After purchasing the project, send requirements so Dejan can start the project.
Delivery time starts when Dejan receives requirements from you.
Dejan works on your project following the steps below.
Revisions may occur after the delivery date.
Baseline & reproduce
I run your current YOLO baseline, verify metrics, and curate a “hard examples” subset.
SAHI parameter sweep
Grid search over tile size / overlap / score & NMS thresholds; measure mAP/F1/latency and pick the best trade-off.