Background

The likelihoods (confidences) produced by SimCCPredictor were not consistent with the expected behavior of SimCC decoding.

• In some experiments, likelihoods exceeded 1, because the predictor was returning raw model outputs (logits) instead of proper probabilities.
• When setting apply_softmax=True, likelihoods were extremely small (~1e-3). This is expected mathematically (softmax distributes probability mass across the entire SimCC space), but not useful as a visibility/confidence measure.
• When setting normalize_outputs=True, most likelihoods were exactly 1, because normalization simply divided by the maximum logit value.

This mismatch originated from how the raw SimCC representation vectors (logits over discretized x/y coordinates) were converted into coordinates and visibility scores.

See the original MMPose predictor implementation for reference.

Changes

This PR brings SimCCPredictor closer to the original MMPose decoding logic:

1. Added parameters

• sigma: variance of the Gaussian used to generate SimCC labels during training. This controls the expected sharpness of the label distribution.
• decode_beta: scaling factor applied together with sigma to logits before softmax. It acts like an inverse temperature in softmax: sharpening peaked distributions (visible keypoints) and flattening diffuse ones (occluded keypoints).

Together, sigma * decode_beta ensures that visibility/confidence is decoded consistently with the label distribution used in training.

2. Likelihood computation

• If normalize_outputs=False (which is the default): the raw logits are scaled by (sigma * decode_beta) before calling get_simcc_maximum with apply_softmax=True. This matches the MMPose decoding step and yields meaningful visibility/confidence scores in [0,1].

3. Default behavior

• apply_softmax now defaults to True (to ensure proper probabilities rather than raw logits are returned).
• normalize_outputs still defaults to False.

Outcome

• Likelihoods are now properly bounded in [0,1].
• The values reflect both the sharpness of the SimCC distribution and the visibility of the keypoint, consistent with the SimCC paper and MMPose implementation.
• Users can still enable normalize_outputs for compatibility with previous behavior, but it is discouraged for real confidence estimation.