Urgent Computer Vision Dataset Audit

We are looking for an experienced Computer Vision Engineer to urgently audit a drive-by dataset collected for a vegetation and roadside-monitoring project. This is the first phase of a larger computer-vision project. We are looking for someone who can start immediately and, if the audit is completed successfully, continue with us on a long-term basis for annotation, model fine-tuning, evaluation and implementation. Project background The dataset contains approximately 1 hour and 40 minutes of drive-by recordings. The available data may include: - RGB images or video frames; - frame and chunk identifiers; - GPS coordinates; - timestamps; - route information; - LiDAR data or LiDAR-derived height statistics; - YOLO-World object-detection output; - SegFormer vegetation-segmentation output; - JSON and GeoJSON files; - previously identified false positives. The purpose of the audit is to determine: - what data is actually available; - how much of the data is usable and unique; - which classes are sufficiently represented; - whether the existing model outputs are useful; - whether GPS and LiDAR can be linked reliably; - what is required for a representative annotation and model-training pilot. Scope of the audit 1. File and dataset inventory Identify and document: - all available files and formats; - total number of videos, chunks and frames; - available RGB, GPS, LiDAR, JSON and GeoJSON data; - missing, corrupted or unreadable files; - links between frames, chunks, routes and sensor data. 2. Image-quality assessment Assess the dataset for: - motion blur; - poor focus; - compression issues; - shadows and backlighting; - distance from relevant objects and vegetation; - visibility of roadsides, vegetation and fixed obstacles; - unusable or low-quality frames. This may be performed using scripts combined with a representative manual review. We do not expect every frame to be reviewed manually. 3. Duplicate and redundancy analysis Determine: - how many identical or near-identical frames are present; - how much redundancy exists within and between chunks; - whether locations or route sections were recorded repeatedly; - how the dataset can be reduced to representative frames; - which frame-selection method should be used instead of automatically selecting every fifth frame. 4. Location-based dataset structure Review: - GPS coordinates and timestamps; - unique locations and routes; - repeated locations and route overlap; - whether training, validation and test data can be separated by location; - how data leakage between near-identical frames and locations can be prevented. 5. Preliminary class distribution Assess whether the data contains usable examples of the following categories. Fixed mowing obstacles - tree trunks; - posts, bollards and lamp posts; - traffic signs; - fences, gates and guardrails; - benches and picnic tables; - utility cabinets, mailboxes and hydrants; - large rocks or boulders; - culverts and bridge structures; - other permanent mowing obstacles. People, animals and moving vehicles are outside the target scope. Vegetation - short or mown grass; - tall roadside vegetation; - reed or watercourse vegetation; - woody vegetation; - cut vegetation or temporary vegetation deposits; - bare or recently mown ground; - other or uncertain vegetation. The audit should indicate: - which classes occur; - estimated example counts per class; - which classes have insufficient data; - which classes should initially be grouped; - which classes are not visually distinguishable with the available image quality. 6. Existing model-output review Review a representative sample of the available: - YOLO-World detections; - SegFormer segmentations; - confidence scores; - false positives; - missed objects or vegetation regions. Determine whether the existing outputs are suitable for automatic pre-annotation and whether correcting them is likely to be faster than annotating from scratch. No model fine-tuning is required during this audit unless separately agreed. 7. GPS and LiDAR assessment Determine: - whether timestamps are consistent; - whether LiDAR data can be linked to the correct frames, chunks and locations; - whether the available calibration and sensor information is sufficient; - whether LiDAR can support vegetation-height analysis; - whether reliable use is possible at pixel, frame, chunk or location level; - which technical limitations or missing information exist. The audit does not include development of a complete RGB–LiDAR fusion pipeline unless separately agreed. 8. Pilot recommendation Based on the audit, provide a clear recommendation for the next pilot phase, including: - recommended number of pilot frames; - frame-selection method; - proposed final class structure; - proposed annotation approach; - location-based training, validation and test split; - estimated annotation effort; - expected technical risks; - additional data that may be required; - estimated scope and cost of the next phase. Required deliverables The final audit must include: 1. documented dataset inventory; 2. counts of available, usable, unusable and duplicate data; 3. overview of unique routes and locations; 4. image-quality findings; 5. preliminary class distribution; 6. review of existing YOLO-World and SegFormer outputs; 7. assessment of GPS and LiDAR usability; 8. recommended frame-selection strategy; 9. recommended pilot dataset; 10. proposed train, validation and test structure; 11. estimated annotation effort; 12. technical risks and missing information; 13. recommended next steps; 14. concise written audit report; 15. all scripts used to produce the audit results. All findings must be reproducible and based on the supplied data. Data-integrity requirement We will independently verify: - file and frame counts; - duplicate percentages; - class counts; - GPS and route findings; - model-output observations; - samples used in the report; - scripts and generated outputs. Do not present invented, synthetic or estimated findings as measured results. Every estimate must be clearly labelled as an estimate and must include the method, sample or assumption on which it is based. Fabricated data, unverifiable metrics, manipulated screenshots or non-reproducible conclusions will result in immediate rejection of the work and termination of the contract. Required experience You should have demonstrable experience with: - computer-vision datasets; - Python; - OpenCV; - YOLO or comparable object-detection models; - semantic segmentation; - SegFormer or comparable segmentation models; - dataset-quality audits; - duplicate and near-duplicate image detection; - train, validation and test dataset design; - GPS or geospatial data; - JSON and GeoJSON. Experience with LiDAR, camera–LiDAR calibration or sensor fusion is strongly preferred. Urgency and availability This is an urgent assignment. Please apply only if you can: - start immediately or very soon; - work directly with the supplied raw data; - communicate progress clearly; - provide reproducible findings; - state your expected completion date before starting. Long-term opportunity This audit is the first part of a broader project. A successful audit may lead to further work involving: - annotation strategy and quality control; - object-detection fine-tuning; - vegetation segmentation; - LiDAR-assisted height analysis; - false-positive reduction; - GPS and GeoJSON output; - evaluation on unseen locations; - production-pipeline development. We are specifically looking for a reliable specialist for a long-term cooperation. Application questions Please answer the following questions: 1. Describe one dataset audit you completed for a computer-vision project. 2. How would you identify duplicate and near-duplicate frames in a long drive-by recording? 3. How would you prevent data leakage between training and test data? 4. What would you check before using LiDAR data together with RGB images? 5. How would you determine whether existing model output is suitable for pre-annotation? 6. How many hours do you estimate for this audit? 7. What fixed price do you propose? 8. When can you start? 9. What exactly will you deliver at the end of the audit?