Designed , built and deployed into production, an End-to-End solution for floor-plan recognition capable to recognize and reproduce all visual aspects and object of a raster floorplan into SVG. Leveraging the spatial complexity in object detection by using oriented detection and rotation equivarient neural networks. ( see thesis abstract in the corresponding section, contact for the full thesis )

• Floor-plan pre-processing : CAD data to raster high quality conversion.

• Data annotation, pre-processing and augmentation for the training.

• Training and deploying state-of-the-art object detection models.

• Built an extensive comparative study of 10+ state-of-the-art Horizontal and Oriented Object Detection.

• Built a Faster R-CNN object detection model with a COCO ResNet101 Transfer Learning backbone to detect logos from input images. This model have been trained on GCP with 4 NVIDIA Tesla k80 GPUs.
• Converted the model weights to TensorRT in order to accelerate the inference when performed on GPU.
• The model predictions are stored in a database on which we applied the Digital Share of Voice metric transformation.
• Built a Dashboard that monitors the results and gets refreshed automatically with each bulk inference done by the model.

• Developed to help Pyronear automatically annotate wildfire images from the web in order to expand the current dataset.
• Uses a pretrained SSD Mobilenet V2 model to make bounding boxes prediction then automatically generate them into XML labels which is the PASCAL VOC 2007 label type
• Model trainined on the wildfire Smoke Dataset released by AI For Mankind with Tensorflow

This project is an implementation of the SLAM (Simultaneous Localization

and Mapping) for a 2 dimensional world:

•  Combines robot sensor measurements and movement to create a map of an environment from only sensor and motion data gathered by a robot over time.