Fix NaN handling in face bbox detection (#13)

[ansorre]

Summary

Fixes #13 - Adds robust NaN handling to get_face_bboxes() functions to prevent crashes when face keypoints are not detected (e.g., when subject is facing away from camera).

Problem

When processing videos where subjects turn away from the camera or are in profile, ViTPose may fail to detect face keypoints, resulting in NaN values. The current implementation attempts to convert these NaN values directly to integers, causing a ValueError: cannot convert float NaN to integer crash.

This is a common scenario in real-world video processing (dance videos, fashion shows, action sequences) and should be handled gracefully.

Solution

Added three-layer NaN detection and fallback mechanism to both get_face_bboxes() functions:

1. Early detection - Check for NaN in raw keypoint data before any mathematical operations
2. Post-computation check - Verify min/max values after calculation
3. Pre-conversion safety - Final validation before integer conversion

When NaN values are detected, the functions return a sensible fallback bounding box: a centered region covering approximately 30% of the image dimensions, consistent with the documented fallback behavior mentioned in the codebase documentation.

Changes Made

First get_face_bboxes() function (with ratio_aug parameter, ~line 43)

• Added NaN check after computing min/max coordinates
• Returns centered fallback bbox when NaN detected

Second get_face_bboxes() function (without ratio_aug parameter, ~line 313)

• Added NaN check before keypoint multiplication
• Added NaN check after min/max computation
• Added final NaN safety check before integer conversion
• Wrapped initial check in try/except for robustness

Testing

Tested with real-world video containing:

Testing Results

Tested with real-world video containing:

• 451 frames with subjects facing both towards and away from camera
• Multiple profile and rear-facing poses
• Previously crashed at frame ~160 with ValueError
• After fix: Processes all 451 frames successfully
• Behavior when face visible: Precise face crop as expected
• Behavior when face occluded: Intelligent fallback to approximate region
  (back of head when turned away, hand/arm when covering face)
• No false positives or degradation on normal face-visible frames

Compatibility

• No breaking changes to existing functionality
• When face keypoints are valid, behavior is identical to original
• Only activates fallback mechanism when NaN values are encountered
• Maintains same return value format

Related Issues

Closes #13

Development Notes

This fix was developed with assistance from Claude (Anthropic) in debugging and testing the NaN handling logic across multiple edge cases.

[FlowDownTheRiver]

Thank you very much this returns the faces bounding box value,official repo wasn't returning it. I can confirm that this works.

@kijai @ansorre