Master's thesis work. Object detection in Sub-T environments towards item search.
This repository hosts the files of the project Convolutional neural networks for object detection in subterranean environments, which aims to explore the capabilities of different state-of-the-art object detectors in the task of detecting some of the DARPA Sub-t Challenge artifacts from image data, then expand and propose a complete perception layer for item search in a mapped environment with a single camera.
Written contents on the official report outweigh those in this repository.
- Research: Gather knowledge on the field of neural networks and object detection. Identify state of the art object detection models and search for implementations backed-up by original research papers that are available for use.
- Data gathering: Produce different datasets to train a model able to identify a specific set of items from pictures.
- Training and benchmarking: Train different neural network models on the gathered data and evaluate and compare their performance on the object detection task.
- Deployment: Build a perception layer suitable for item search around an object detection approach.
- Model benchmarking:
- Benchmarking results:
- See project report or showcase below
- Full COCO tables available at result-files
- Generated data:
- Real samples: PPU-6 dataset
- Synthetic samples: Unity-6-1000 dataset
- Trained models: available here
- Benchmarking results:
- Perception layer for item search
- Find source code and instructions under ./deploy-remote/perception-layer-final
