Two models that find 10 kinds of objects in drone photos, such as pedestrians, vehicles and bicycles, run side by side on the same photo. Both are the same model, trained only at different input sizes.
The model runs in the browser; uploaded images are not saved.
An object detection demo that runs entirely in the browser and finds 10 object types such as pedestrians, cars and bicycles in drone photos. Two versions of the same model can be compared side by side on the same photo.
Most objects in drone photos are tiny. A pedestrian about 20 pixels tall in a 1360-pixel-wide frame shrinks to 9 pixels when the image is scaled down to 640 pixels for the model. At the standard input size, accuracy on small objects stays low (AP 0.111).
The same YOLO11s model was trained on VisDrone at two input sizes, 640 and 1280 pixels. The 1280 input raises small-object AP from 0.111 to 0.223. The cost is a longer run time (65 ms instead of 23 ms on WebGPU). Both versions are exported to ONNX and run side by side in the browser.
Same model, same data. Ultralytics' own tool gives mAP50 0.380 for Model-1, the standard COCO tool pycocotools 0.347. The two match boxes in a different order and compute AP differently. Every score on this page comes from detection-eval, written from scratch.
The photo is shrunk before it reaches the model. Averaging every pixel a new pixel replaces (antialias), instead of reading only the nearest few, raises Model-1 from 0.347 to 0.355. The gain is spread over all object sizes.
Running Model-1 on overlapping 640 × 640 tiles of the full-resolution photo (SAHI-style) raises mAP50 from 0.355 to 0.425. The key is dropping the partial boxes of objects cut by a tile border. The cost is about 6 model runs per photo. Model-2, trained at 1280, is both better (0.505) and faster without tiles.
On a desktop GPU (RTX 3060 Laptop) the model was converted to TensorRT FP16 and the resize moved to the GPU. Model-1's total time per photo went from 15.2 ms to 5.6 ms, with mAP50 down by only 0.0013.
11/01/2026
They find 10 kinds of objects: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus and motor. The models were trained on the VisDrone dataset of drone images.
Both are the same YOLO11s model. Model-1 looks at the image scaled to 640 pixels, Model-2 at 1280 pixels. Model-2 is better on small objects but runs slower. On the VisDrone validation set, mAP50 rises from 0.347 to 0.503.
Nowhere. The image is opened in the browser and the model runs directly on the device. It is not sent to a server or saved.
Yes. JPEG, PNG, WebP, AVIF, BMP and GIF images can be uploaded. Since these images have no labels, correct and false detections are not separated; the objects found are only marked by class.