Estimation des disparités par adaptation de domaine pour la détection d’objets 3D à partir d’images stéréoscopiques

Abstract

Today’s world is confronted to several devastating consequences due to pollution. Increasingly, people want to adopt an eco-friendly behavior and find a substitution to the car is the first step to reduce pollution. And the bycicle is the means of transport the most voted in. However, the coexistence between bicycles and cars is difficult and causes many serious accidents every year, which requires effective solutions to ensure the safety of cyclists. In this context of road safety, we have developed a system using stereovision and a 3D detection method in deep learning, with the aim of estimating the safe distance between a bike and a vehicle when passing. However, the actuals stereo-image based 3D detection methods have limited performance, mostly due to the disparity estimation methods used to produce 3D representation of the scene before detection. Indeed, obtaining real disparity map densly annoted is a tedious task, and the few dataset available does not allow an efficient training of the model. In this thesis, we will focus on the design of a fast disparity model, inpired by the 2D network DispNetC, whom we will seek to improve its accuracy and generalization. Moreover, in order to solve the problem of lack of real labeled data, the model will be adapted to the real domain in a unsupervied manner by the adversarial learning principle, allowing to reduce the gap between synthetic (annotated) and real (without annotation) domains. The CycleGAN network will translate synthetic learning data into the real domain, and a feature discriminator will make the internal representations of the network invariant to the domain. The proposed learning strategy will increase the robustness of the model facing domain shift, and improve its performance on the real domain in the unsupervised case. The final model will be combined with a image-stereo based 3D detection method to measure its ability to produce 3D informations for detection. It will be evaluated on the KITTI dataset for 3D detection and the results of the experiments will be validated by comparison with literature methods.

Date	21 Feb 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Marco Pedersoli (Supervisor)