V2-SAM Ego2Exo checkpoints

This repository preserves the file layout of wangzeze/V2-SAM at revision 50fd5a9a7e67d3fdaadab1cd0726b82896f89e02. Only the two Ego2Exo checkpoints were replaced with the best checkpoints from the audited reproduction. The DINOv3 asset, dataset archive, and both Exo2Ego checkpoints are byte-for-byte copies from that pinned public model repository.

The matching NewMatcher source is jaychempan/V2-SAM-O@travor at commit bd7ba86314b8a2916e4af0d55eba1ed1b37c271e.

Ego2Exo checkpoints

File	Model / checkpoint selection	Object-instance mean IoU / Dice	Frame/pair-macro IoU / Dice
fusion_ego2exo_full.pth	Strict NewMatcher Fresh24, epoch 16 (best of epochs 2–24)	0.405197334183 / 0.462906654638	0.450898164298 / 0.509557001338
vp_ego2exo_full.pth	Public Visual, resumable from-scratch 24-epoch run, iter_27600 (best retained checkpoint)	0.3683 / 0.4305	not recomputed

Metric aggregation

• Object-instance mean gives equal weight to all 100,223 object–frame instances from 40,517 test frame pairs. This is the aggregation used by the complete-object audit and is the 0.405197 NewMatcher value.
• Frame/pair macro first averages all objects inside each frame pair and then gives equal weight to each of the 40,517 pairs. This is the higher 0.450898 NewMatcher value. It is also referred to as the author's frame-level aggregation in the experiment records.
• These two numbers describe the same raw predictions under different aggregation weights and must not be compared as if they were one metric.

Training contracts

Strict NewMatcher Fresh24

• Fresh training: load_from=None, resume=False; only the SAM2 and DINOv3 foundation weights were loaded.
• Optimizer: AdamW, learning rate 4e-5, effective batch 64 (batch_size=16, accumulative_counts=4).
• Schedule: Linear warm-up from epoch 0 to 1.2, then cosine decay through epoch 24.
• The 24-epoch run was evaluated at epochs 2/6/8/10/12/14/16/18/20/22/24; epoch 16 was the final best checkpoint.
• Checkpoint SHA256: 7a75826ffb7f065e7db9c20322567eb6058c9697e079180f77cda9ad4097d5b3.

Public Visual resumable from-scratch

• Fresh 24-epoch training using the public Visual architecture.
• Optimizer: AdamW, learning rate 4e-5; Linear warm-up to epoch 1.2, then cosine decay through epoch 24.
• iter_27600 was selected by full-test evaluation; later retained/final checkpoints did not improve it.
• Checkpoint SHA256: bc94b285f5421359702b2c42ef2aec83ffee6787b301dea5821628ccdca8d0c4.

The package keeps the exact top-level file layout of the pinned public model repository. Large files are tracked with Git LFS according to its unchanged .gitattributes.