Internship - Event-Based Drone Detection Annotation Pipeline

Location: Paris (75011)Contract: InternshipStart date: Between January and March 2027Why PROPHESEE?Prophesee is pioneering a new generation of machine vision inspired by the human eye and brain. Its breakthrough event-based vision technology enables machines to see and respond to changes with unprecedented speed and efficiency. It serves demanding markets including industrial automation, aerospace and defense, autonomous systems, and IoT.This 6-month end-of-studies internship is based in Paris (75012), in the Software team. Internship topic : Semi-Automatic Annotation Pipeline for Event-Based Drone DetectionTraining drone detection models requires large, accurately annotated datasets. Manual annotation is time-consuming, especially for small, fast-moving drones recorded with event cameras.Recent detection and tracking models such as SAM 2.1 and ETAP can speed up annotation by detecting drones and propagating bounding boxes through time. However, tracking can fail or drift, automatically generated labels can contain errors, and models pretrained on standard images need to be adapted to event-based data.The goal of this internship is to build a robust semi-automatic annotation pipeline that generates, validates and corrects drone annotations while minimizing manual work.Main ResponsibilitiesThe intern will contribute to the following activities:Benchmark automatic annotation methods (our in-house drone detector, SAM 2.1, ETAP and other detection/tracking approaches) for generating and propagating drone bounding boxes.Adapt pretrained RGB models to event-camera data: event-to-frame representations, event-to-video reconstruction, domain adaptation and lightweight fine-tuning of selected model components (e.g. input layers or decoders).Improve tracking robustness: detector-guided initialization and re-initialization, forward/backward tracking, temporal consistency, and automatic detection and correction of tracking failures.Improve pseudo-label reliability by combining several sources of evidence: multi-model ensembles, test-time augmentation and consistency checks, temporal validation, and, when available, geometric information (GPS, camera orientation, calibration).Design confidence and uncertainty metrics to assess annotation quality and detect errors.Develop annotation cleaning and human-in-the-loop strategies: flag suspicious annotations, correct them when possible, and prioritize uncertain frames for human review (including active learning approaches).Integrate the tool with our dataset versioning infrastructure (navigate, download and update datasets).Evaluate the pipeline in terms of annotation quality, robustness, computational cost and reduction of manual labeling effort, and document results and recommendations.Expected ProfileWe are looking for a final-year student (engineering school or Master’s degree) for an end-of-studies internship in computer vision, machine learning or a related field.The ideal candidate has:Strong knowledge of deep learning for computer vision, especially object detection and/or tracking.Good programming skills in Python.Experience with PyTorch.Experience working with Linux and Git.Fluent English.The following are a plus:Experience with foundation models (e.g. SAM), pseudo-labeling, active learning or uncertainty estimation.Familiarity with event cameras or neuromorphic vision.Knowledge of camera calibration and multi-view geometry.Experience with data pipelines or dataset versioning tools.MindsetStrong problem-solving skills, strong analytical skills. Flexible to dynamic environments and fast changing technologies. Passionate about technology. This person must work well with other engineers in a team environment. Good sense of autonomy. Must be pragmatic and self-motivated to complete a task even if it is outside of just the “well known” realm : “Can Do Attitude”.What You Will LearnDuring this internship, you will gain experience in:Event-based vision and neuromorphic perception.State-of-the-art detection, tracking and foundation models applied to a new type of sensor.Domain adaptation from frame-based to event-based data.Pseudo-labeling, confidence estimation and human-in-the-loop annotation.Building production-grade data tools in a deep-tech environment.