Make Class-Agnostic 3D Segmentation Efficient with 3DIML

Under uniform Mask2Former frontends, we benchmark against Panoptic and Contrastive Lifting while evaluating 3DIML on the Replica-vMap and ScanNet datasets. While cutting neural field training iterations by 25×, 3DIML shows comparable accuracy using Scene Level Panoptic Quality and mIoU measures. Comparing 3DIML to Panoptic (5.7h) and Contrastive Lifting (3.5h), the former takes less than 20 minutes to finish scans on a single RTX 3090.