ФранцияResearch Intern
вчера
ФранцияTraineeHybrid
background subtractionobject detectionsegmentationdinogrounding dinosam 2
Research internship focused on developing video processing and object detection methods for urban scene monitoring under computational constraints.
Требования
- Final-year student in an M2 or engineering school, specialised in computer vision, machine learning or applied mathematics
- Fluent in Python and PyTorch, and comfortable processing video streams
- For Subject A: knowledge of object detection, multi-object tracking or visual representation learning. For Subject B: knowledge of generative models (GAN, diffusion, inpainting), detection, tracking or re-identification, and an interest in personal data protection
- Technical English good enough to read and synthesize research papers
- Rigorous with experiments and able to move a research question forward on your own
- At , we care about precision, ownership and getting things done. We're looking for people who are genuinely interested in what public-sector work involves.
Условия
- 🏡 Hybrid work 🏝 Mandatory internship agreement from your school (1 day off per month) 💻 MacBook ou PC 🍜 60% coverage of meal vouchers worth €9 per day worked 🚃 50% Navigo pass reimbursement 💼 Offices located in central Paris (9th arrondissement) ☀️ Annual offsite with the whole team and plenty of company events
- If you've read this far, you're probably very interested in the role and/or . We'd like to hear from you even if you feel you don't check every box.
Другое
- 's cameras stay on the same streets for months and process video in the field, under strict compute and energy constraints. Two research questions come directly from that setup. How do you understand what changed in a scene over hours or days, when classic methods only compare two frames? And how do you use urban video to train models and deliver clips to local authorities, while protecting the people who appear in it without having anything to do with the event?
- 250+ local authorities already rely on to keep their streets clean and safe. Each of these questions affects what our cameras can detect and how we handle the data behind it.
- We are opening two end-of-studies research internships in our R&D team, one per question. Same internship, same setup, two subjects. Each is built around a research question, a hypothesis to test, public benchmarks and long real-world sequences, and is designed to support your master's thesis. The internship lasts 6 months, with the possibility of continuing as a CIFRE PhD.
- Implement pixel-level baselines (background subtraction, learned change detection on image pairs) and low frame rate tracking, and quantify where they fail
- Design an object-centric scene memory: detect objects with open-vocabulary detectors (Grounding DINO, OWLv2), segment them (SAM 2), describe them with lighting-robust features (DINOv2) and their presence history, and assign each object a state updated over time: permanent, temporary, new, removed
- Target cases from 's field use: illegal dumping, abandoned luggage, vehicles parked for days, damaged or missing street furniture
- Characterise failure modes by observation rate, history length, day/night and weather, object shape and evolution (a pile of waste that grows, a tarp moving in the wind), scene density and occlusions
- Evaluate on events rather than frames: precision and recall of appearances and disappearances, detection delay, false alerts per camera per day, on PETS 2006, i-LIDS / AVSS 2007, ABODA, CDnet 2014 and sequences
- Build detect-then-blur baselines (RetinaFace, EgoBlur, 's in-house plate detector), measure missed detections first (small faces, night, motion blur), then quantify the recall gained by propagating anonymisation through tracking
- Compare generative replacement methods: DeepPrivacy2 full-body GAN, diffusion inpainting for faces, full-body diffusion and video regeneration from structural cues (pose, segmentation, depth), plus synthetic plates with known text
- Work on temporal consistency (one stable synthetic identity per track, no flicker) and on selective anonymisation for clips delivered to local authorities: everyone except the people involved in the event
- Measure protection (residual leaks per 1,000 frames, ArcFace and person re-ID attacks, OCR on anonymised plates) against utility (detection and tracking performance of models trained on raw vs anonymised video, event readability for an operator), within the frame of GDPR and the CNIL's 2022 position on augmented cameras
- Evaluate on MOT17, MOT20, Cityscapes, BDD100K and sequences
- Start with a literature review, then build test sets from public benchmarks and long sequences, part of which you will annotate during the internship
- Define the evaluation protocol and metrics before running the main experiments
- Deliver a model or pipeline evaluated on public and data, a documented benchmark, clean code, integration recommendations and your master's thesis