6-month research internship on contrastive learning for images. Here, my work has focused on using foundation models (like SAM) to select better views for contrastive learning (which is usually done randomly). Accepted at NeurIPS 2023 workshop [1], and under review at CVPR 2024 [2]. I also invented a method for batch image annotation (pending patent) [3].

[1] B. Missaoui, C. Yuan, “SAMCLR: Contrastive pre-training on complex scenes using SAM for view sampling” (Accepted at NeurIPS'23 workshop)

[2] B. Missaoui, C. Yuan, T. Ngo “Towards disambiguated self-supervised learning” (Under review at CVPR 2024)

[3] T. Ngo, B. Missaoui et al., Automated Image Annotation Method And System (EU patent)

Research project on 3D reconstruction with Neural Radiance Fields (NeRFs), which we extended to be able to generate any wavelength (not just RGB). This enables all new use cases in biology and recycling, where different plants or materials exhibit different properties depending on the wavelength they are viewed from. Submitted to ICCV 2023 [4].

[4] G. Chen, H. Muriki, B. Missaoui, al., “Hyper-NeRF: Hyperspectral Neural Radiance Fields with Continuous Radiance and Transparency Spectra” (Submitted to ICCV 2023)

Developed a solution fusing camera, LiDAR and IMU for projecting landmarks from 2D maps to the vehicle’s camera frame, enabling to get high-quality-pseudo labeled training data for object detection models.

The method was accepted at IEEE IV 2023 [5], the most prestigious conference on intelligent vehicles.

[5] B. Missaoui, M. Noizet, P. Xu, “Map-aided annotation for pole base detection” (Accepted at IV 2023)

6 months research internship on Deep Unsupervised Anomaly Detection. Goal: Improve defect detection on industrial parts.

• Co-authored a heuristic to remove defective parts in a fully unlabelled dataset with 90+ AUC.
• Fine-tuned pre-trained networks with synthetic defects to improve defect detection AUC from 91.6 to 96.1 points.