This project was developed as part of the AI Grand Challenge 2024–25 (https://www.ai-grand-challenge.in/), jointly organized by NTRO (National Technical Research Organization) and IIT Delhi(NCIIPC). The challenge focused on large-scale AI-driven geospatial analytics, requiring automated annotation, object discovery, and intelligent search across 3-meter resolution satellite imagery.
GeoAccel-AI (this repository) represents my final end-to-end solution — covering prototype generation, feature distillation, GPU-optimized auto-annotation, segmentation-guided refinement, hybrid DINO + SAM2 embeddings, YOLO training, and competition-ready inference pipelines.
The complete codebase used for official submission
Final results on the evaluation dataset
GPU-optimized workflows, profiling reports, and visual search pipelines
All tools built to accelerate annotation workflows during the challenge
Mapped Score Acheived against the cut-off of 1.100
Accelerating Geospatial Intelligence through distillation, segmentation, and proprietary embeddings.
Developed under the SvarAikyam AI / AI Fusion initiative, this project integrates deep visual distillation models, segmentation refinement, and GPU-accelerated feature search to automate large-scale annotation and object discovery in high-resolution (3m) satellite imagery.
GeoAccel-AI enables automated detection, labeling, and visual search across multi-band satellite datasets — empowering geospatial workflows through AI-driven similarity and segmentation refinement.
The visual search can identify and classify diverse infrastructure and terrain categories with minimal human supervision.
Supported object classes:
Brick Kiln | STP | Solar Panel | Sheds | Metro Shed | Play Ground | Pond-1 | Pond-2
| Stage | Description |
|---|---|
| Prototype Creation | Extracts proprietary feature embeddings from annotated .json polygons (ground truth). |
| Batch Prototype Builder | Builds class prototypes across all TIF + JSON pairs. |
| Auto-Annotation (GPU Optimized) | Performs multi-scale window detection using embedding similarity search and segmentation-based refinement for detected regions. |
| Batch Detection | Executes large-scale annotation runs across hundreds of satellite tiles, generating object-level metadata. |
| Model Training | YOLOv8-based detection fine-tuned on discovered regions; used as a region proposal network (RPN) for refinement. |
| Final Visual Search | Embedding-based similarity retrieval to identify query objects in unseen images. |
| Interactive Review UI | Provides OpenCV-based annotation validation, class editing, and YOLO export tools. |
|
Below are the metrics for a representative YOLOv8 model trained for 200 epochs.
Since the extracted embeddings contain a degree of variability, moderate precision is expected — yet sufficient for rapid bootstrapping of geospatial datasets.
Figure 1: YOLOv8 training loss curves over 200 epochs.
Integrated with Nsight Systems, Nsight Compute, and Torch Profiler for kernel-level insights, enabling precise GPU workload tuning. Includes experimental support for Agentic GPU Optimization, orchestrating autotuning of CUDA kernels and Triton ops for real-time efficiency gains on *RTX 3060.
Figure 2: Initial Profiling Report on RTX3060 (12GB).
Figure 3: Optimized Profiling Report on RTX3060 (12GB).
Figure 4: Initial Annotation for Training.
Figure 5: Final Result — Detected Known Objects in Unseen Image.
Figure 6: Final Result — Detected Unseen Objects in Unseen Image.
GeoAccel-AI represents a unified, GPU-optimized geospatial AI framework that fuses:
- Visual distillation for semantic and texture representation
- Segmentation refinement for pixel-level precision
- Agentic profiling & kernel discovery for dynamic GPU efficiency
It serves as a bridge between remote sensing, AI-driven automation, and edge-to-cloud scalability, enabling faster Earth observation analytics and infrastructure monitoring.
@misc{geoaccel_ai_2025,
title = {GeoAccel-AI: GPU-Optimized Auto-Annotation for Satellite Imagery},
author = {Atul Vaish},
year = {2025},
url = {https://github.com/intelav/GeoAccel-AI}
}
© 2025 SvarAikyam AI | AI Fusion — Applied Research in GPU Optimization & Geospatial Intelligence