Crowd counting

Recent advances in deep learning have show great potential for crowd counting from 2D images. While several solutions require dedicated cameras (2d or 3d) for high precision counting, recent methods leverage existing surveillance cameras and achieving decent performance around 90% accuracy. The severe occlusion in crowded spaces remains challenging but promising performance are obtained in somewhat controlled indoor environement. Density-based approaches as well as head detection approches have been evaluated on high view point scenes in airports and close view point in buses.

OCR on raw leather markings

Traditional and off-the-shelf OCR solutions could not read leather markings correctly due to various text orientations and special image texture. Hence, a custom OCR solution was developed and patented. As the Lead Tech in this project, I was designated as one of the inventors.

For more information refer to Neovision project page or directly to CTC’s project page

Ear landmark detection

In order to design custom eartips, 3d scans of the ear and its canal are required. A solution based on CNNs has been developped in order to track ear landmarks during smartphone scaning. Inspired by facial landmarks research, a deep learning based solution was trained on a proprietary database to detect ear landmarks. The solution achieved high accuracy and has been deployed successfully.

Electrical symbol recognition from schema

Digitizing Single Line Diagrams (SLD) is a long and expensive process due to the large diversity of symbols and the high number of occurences. Recent advances in object detection and fine-grained image classification offer good solutions to detect and recognize electrical symbols from schemas automatically. A proprietary database was constructed to train and evaluate state of the art solutions for spotting and then recognizing each symbol.

Food volume estimation

Food volume estimation is an important challenge with many applications ranging from measuring calories in a diet management program to estimating the cooking time depending on the nature and thickness of the ingredients when following a recipe.

A study was conducted to experiment vision-based solutions and a prototype was developped in the form of an android app. From multiple pictures of the ingredients, the app is able to recover the thickness and surface of each ingredient with good precision.

Invoice parsing

Accounting entry of invoices is a tedious task and many solutions exist on the market to help accelerate the process. Traditional image processing combined with rule-based algorithms do not cope well with the growing diversity of invoice templates. Hence, deep learning based solutions with continuous training are now the leading tehnologies as they can adapt to new templates provided that enough examplars are seen. In this work, a generic state of the art OCR service has been deployed along with a custom trained key information retrieval neural network model to infer the different accounting fields on the invoice.

Construction vehicle detection

Time-lapses of construction sites provide a great way to monitor and supervise construction activities. Construction vehicule presence is a good indicator of the building stage and having statistics over time provide valuable insight for monitoring. While construction vehicule have only a few classes (dump truck, compactor, excavator, loader, mobile crane), their appearance changes a lot due to their articulated structure and many disctractors are present on the scene. A performance study as conducted to evaluate state of the art object detection solutions on this task.