Senior Machine Learning & Computer Vision Engineer for SOTA Challenge

Overview I'm looking for a senior machine learning / computer vision engineer with genuine medical imaging experience to take two existing dental radiograph models and re-architect them to beat their current baselines and prior published SOTA on the same 31-class dental dataset. This is not a "run the training script" job. I already have working baselines. What I need is someone who can introduce novel architectural modifications, loss/matching changes, or training strategies and demonstrate a measurable, defensible improvement over both the baseline and the best previously reported numbers on this dataset. The work splits cleanly into two independent projects. You can take one or both. The Dataset (shared across both projects) A dental radiograph dataset with 31 classes (dental pathology + restorations + anatomy), including Caries, Crown, Filling, Implant, Root Canal Treatment, Impacted tooth, Periapical lesion, Bone Loss, Mandibular Canal, Maxillary sinus, and rarer long-tail classes (TAD, Cyst, Root resorption, etc.). Train: 9,481 images / 94,794 annotations Valid: 2,871 images / 27,141 annotations Test: 1,580 images / 14,957 annotations Important characteristic: the class distribution is severely long-tailed. Some classes have 8,000–33,000 instances (Filling, Impacted tooth, Root Canal Treatment) while others have single-digit instance counts (Fracture teeth = 9, TAD = 4, Bone defect = 1). Handling this imbalance well is central to the challenge. Data is available in both YOLO polygon format and COCO JSON. Project 1 — Detection (DINO-DETR / DETR-family): $400 Starting point: the official DINO-DETR ResNet-50 4-scale COCO checkpoint (checkpoint0033_4scale.pth). To be clear about what this checkpoint is, so we're on the same page: it is the stock COCO release used only as initialization weights (num_queries=900, backbone=ResNet-50, 91-class COCO class embedding, dataset_file=coco). Out of the box it predicts COCO categories — person, car, dog. It has no dental knowledge and must be fine-tuned on the dental COCO JSONs before it's useful. Your task: Fine-tune / adapt DINO-DETR (or a justified DETR-family variant) to the 31 dental classes. Introduce novel modifications — this is the core deliverable. Examples of acceptable directions (your call, justify it): improved query design, denoising/matching strategy changes, multi-scale/backbone changes, long-tail-aware losses, better handling of tiny/low-contrast lesions like caries. Beat the fine-tuned baseline and prior reported SOTA on this dataset (COCO mAP, and per-class AP for clinically important classes). Project 2 — Instance Segmentation (YOLOv8-seg): $400 Starting point: a trained YOLOv8-seg model (best.pt). For clarity: this is a segmentation model (task=segment, Segment head with proto + mask coefficients, C2f/SPPF backbone, imgsz=640, trained ~50 epochs). The YOLO label files are polygon masks (class + polygon points), not bounding boxes. Note the COCO JSON export for this set currently has segmentation: [] empty (boxes only) — so mask supervision comes from the YOLO polygon labels, and part of the job may involve reconciling the two formats cleanly. Your task: Improve the segmentation architecture/training with novel changes (e.g., mask head refinements, boundary-aware or long-tail-aware losses, augmentation strategy, backbone/neck changes, or a stronger seg architecture entirely if you can justify the switch). Beat the baseline and prior SOTA on segmentation metrics (mask mAP, plus per-class where it matters). Deliverables (per project) Modified, well-documented training/inference code (reproducible, seeded). Clear ablation study isolating the contribution of each change — I want to see why it improved, not just that a number went up. Final metrics vs. baseline vs. prior SOTA, on the held-out test split (no test leakage — this is a hard requirement). Trained weights + a short technical write-up explaining the novel contribution. Required Skills & Qualifications Strong track record in medical / clinical imaging AI (dental, radiology, or comparable). Deep hands-on expertise with DETR-family detectors (DINO-DETR, Deformable DETR) and YOLO / instance segmentation. Proven ability to modify architectures, losses, and matching — not just run existing repos. Solid grasp of long-tailed / class-imbalanced detection and segmentation. Fluent with PyTorch, Ultralytics, COCO eval, and rigorous experimental methodology (ablations, proper splits). Bonus: prior publications or documented SOTA results on public benchmarks.