Superanimal humanbody: Use filtered detector

[maximpavliv]

Summary

This Pull Request introduces a new FilteredDetector class, which wraps a generic object detector (typically trained on COCO) and filters its outputs to retain only detections for a specified class_id.

It also adds a utility function, get_filtered_coco_detector_inference_runner(), which returns an inference runner using a FilteredDetector wrapping one of the following torchvision models:

• fasterrcnn_resnet50_fpn
• fasterrcnn_resnet50_fpn_v2
• fasterrcnn_mobilenet_v3_large_fpn

Main changes

• Replaces the existing superanimal_humanbody inference pipeline with a FilteredDetector-based DetectorInferenceRunner.

• Solves multiple issues in the current superanimal_humanbody integration:

  • Fixes temporal shift in predictions
    

  • Prevents CTD inference pipeline crashes (which solves this comment)

  • Resolves detector training bug where loss remained at 0.0

  • Removes unnecessary debug prints

Structural improvements

• Integrates superanimal_humanbody more cleanly into the inference flow: it is now handled via deeplabcut.pose_estimation_pytorch.modelzoo.inference._video_inference_superanimal(), like other PyTorch-based superanimal models.

• Exposes FilteredDetector as a reusable utility for users who want to focus on specific class_ids.

• Improves separation of concerns:

  • Preprocessor handles input data preparation
  • Postprocessor handles filtering and final outputs

Related PR

This PR is paired with DLClibrary PR #42, which extends dlclibrary.download_huggingface_model() to rename the humanbody pose estimation snapshot upon download.