This is a finetuned version of SpatialLM1.1-Qwen-0.5B on ARKitScenes, using a new set of object categories and enable random, gravity-aligned scene orientations. During inference, only the z-axis needs to be kept as the up axis; the x and y axes do not need to be aligned.

Here are some example results comparing ground truth (GT) oriented object bounding boxes with predictions (Pred) from the finetuned SpatialLM1.1-Qwen-0.5B model on the ARKitScenes dataset.