Refactor & Enhance: Inference Pipeline, ONNX Export, and Core Modules

[dnth]

This PR introduces significant enhancements and refactoring across the codebase, focusing on improving the inference pipeline, streamlining ONNX export, enhancing core module efficiency, and updating documentation and setup.

Key Changes:

• Inference & Visualization:
  • Refactored scripts/live_inference.py for improved performance, readability, and added support for video file input alongside existing webcam and image options.
  • Enhanced the draw_boxes function for clearer object detection visualization using OpenCV, including accurate scaling/padding adjustments for bounding boxes.
  • Improved the Gradio demo (gradio_demo.py) visualization logic.
  • Added FPS, provider, and resolution display overlays to live inference outputs.
• ONNX Export:
  • Introduced a PreprocessingModule to optionally embed image preprocessing steps (resizing, color conversion, normalization) directly into the exported ONNX graph.
  • Updated the export.ipynb notebook with detailed instructions and options for ONNX model export.
• Core Module Refactoring & Fixes:
  • Optimized tensor concatenation in box_ops.py (box_xyxy_to_cxcywh) and dfine_utils.py (distance2bbox) using torch.cat.
  • Improved tensor flattening in hybrid_encoder.py using reshape.
  • Updated the PostProcessor in postprocess.py to handle orig_target_sizes as an optional input.
  • Refactored components within the DFINE Transformer (dfine_transformer.py).
  • Corrected label format handling in coco_eval.py (CocoEvaluator).
• Training & Configuration:
  • Updated dataset paths and configurations in train.py for the Rock Paper Scissors dataset.
  • Adjusted the flat_epoch value in deim_hgnetv2_n_coco.yml.