CAMDet — Camera-Agnostic Metric 3D Detection

CAMDet works out where objects are in 3D — their size, distance and position — from a single ordinary photo, with no depth sensor. The catch it tackles: most single-image 3D detectors are quietly tuned to one specific camera, so pointing a different camera at the same scene drops their accuracy, because every lens “sees” the world a little differently.

It's a single transformer-based detector that predicts each object's 2D box, its distance, and its full 3D box together from one shared set of queries — reusing the predicted depth as the object's 3D centre. The key move is training the model to be robust to the camera itself, so one set of weights generalises to sensors it never saw during training instead of overfitting to the camera it was trained on.

CAMDet beats strong prior methods at every model size and runs in real time — up to about 128 frames per second at its smallest scale. On cameras deliberately held out of training it outperforms a standard baseline by a wide margin, and the exact same exported model runs unchanged all the way from a desktop GPU down to a battery-powered edge device.