Benchmark

HardMatch

A benchmark of difficult image pairs for evaluating pose-estimation failure detection in feed-forward reconstruction models, containing 1,000 deliberately selected difficult image pairs.

Timeline (last 90 days)

Oct 27

The research team built the HardMatch benchmark, containing 1,000 deliberately selected hard image pairs.

Unverified50%

Oct 27

On the HardMatch benchmark, VGGT fails on 35% of the image pairs.

Unverified50%

All Facts (2)

Unverified

The research team built the HardMatch benchmark, containing 1,000 deliberately selected hard image pairs.

50%Unverified

On the HardMatch benchmark, VGGT fails on 35% of the image pairs.

50%

Source Articles

VGGT's Self-Awareness: Frozen Features Predict Pose Failures, ROC-AUC 93.5%Oct 27