FlexGlue: early stopping of non-covisible pairs - #4916

Stacked on #4915 (GPU feature cache) and #4908.

Summary

The default FlexGlue matcher models now stop early on pairs that are confidently not covisible:

• After transformer layers 2, 4, 6 and 8 (of 9), a pair is rejected without running the remaining layers and the readout if its covisibility score is below min_covisibility * (1 - early_stop_decay)^(9 - layer).
• The new option FlexGlueMatching.early_stop_decay defaults to 0.5. 1 disables it, as does min_covisibility = 0.
• early_stop_decay is a required input of the matcher models, as in the new default models.

The new models are exported by cvg/FlexGlue#7 (stacked on cvg/FlexGlue#6, fused GELU):

• The checks are ONNX If nodes built on the exported graph. The GELUs use ONNX Runtime's fused operator.
• The models are flexglue_matcher_early_stop.onnx and flexglue_matcher_early_stop_fp16.onnx. The descriptor is unchanged.
• With early_stop_decay = 1, they give the same matches as before. fp32 on CPU matches the PyTorch reference exactly; fp16 differs by at most 2 matches per pair.

The models are in the FlexGlue v0.1 release and are exported by cvg/FlexGlue#7.

Do the ONNX If nodes actually skip the remaining blocks? Yes.

fp16, N = 2048, ONNX Runtime 1.23, A100, device I/O:

• A pair that stops after layer 1 takes 3.5 ms instead of 32 ms. After layer 2, it takes 6.8 ms.
• Each layer costs about 3.3 ms, and nothing after the stop runs.
• Cost on pairs that run all layers:
  • GPU time: +0.3 ms per pair for the 4 checks (the covisibility heads).
  • Wall time: about +0.2 ms per check, because ONNX Runtime needs each condition on the host and synchronizes.
  • The fused GELU of the new models partly offsets this. In COLMAP, pairs that never stop take +0.6 ms at 2k (54.5 vs. 53.8 ms/pair for FlexGlue: cache the dense features on the GPU #4915, +1%) and +0.3 ms at 8k.
• GPU memory: the subgraphs add about 0.55 GB of arena memory at 2k (2.2 → 2.8 GB) and practically none at 8k (5.8 GB).

How much do we gain?

Setup:

• 3 ETH3D DSLR scenes, 2132 exhaustive pairs, on an idle A100.
• Total matching time / pairs, including image decoding and the descriptor.
• For each scene, the fastest of 3 (2k) or 2 (8k) repeats.
• "No early stop" is early_stop_decay=1.

The gain is larger at 8k, because the readout that a stop skips costs more with more keypoints.

Over all exhaustive pairs (per-layer covisibility scores of every pair, with the time per layer measured above):

• The gain depends on the share of non-covisible pairs. In IMC's covisible pairs, the scores only drop in the late layers.
• Lower decays stop more often but start rejecting true pairs:
  • decay 0.3: −18% to −23% matcher time on ETH3D, losing 1.8% of the kept pairs;
  • decay 0.2: about −30%, losing 4–5%.
  • decay 0.5 is a safe default.

The whole stack (#4915 + this PR), with defaults

Same benchmark:

With min_covisibility=0.8, the stack takes 50.3 ms/pair at 2k (−10.1%) and 58.0 ms/pair at 8k (−15.3%).

Tests: 29/29 feature/, controllers/ and util/onnx tests pass, including the new FlexGlueTest.EarlyStopping.