MAP: Unleashing Hybrid Mamba-Transformer Vision Backbone's Potential with Masked Autoregressive Pretraining

Benchmark Model Rank Results
3d-part-segmentation-on-shapenet-partMamba3d+MAP (extra data)–Instance Average IoU: 86.3Class Average IoU: 84.7
3d-point-cloud-classification-on-modelnet40HybridMT3D+MAP (extra data)–Overall Accuracy: 95.9
image-classification-on-imagenetHybridMH-L + MAP–Top 1 Accuracy: 86.2%