[FEATURE] Add a Standalone CLI Inference Script With Bounding-Box Visualization

[Chiragjjjks]

Is your feature request related to a problem? Please describe.
The project currently does not provide a unified, ready-to-use inference script. Users who want to run inference on trained EfficientDet models must write custom scripts, modify internal code, or rely on notebooks.

This becomes a barrier for:

• Beginners exploring the repo
• Users testing trained checkpoints
• People integrating inference into pipelines
• Contributors reproducing results

Additionally, the repo lacks a built-in utility for drawing bounding boxes, labels, and confidence scores on images or videos, something essential for interpreting predictions.

Describe the solution you'd like
I propose adding a standalone command-line inference script, e.g., inference.py, that:

✔ Runs entirely from the terminal

Using argparse, allowing users to run inference with a single command.

✔ Accepts flexible inputs

• Path to model checkpoint (.pth or .tar)
• Path to a single image, directory of images, or video
• Output directory
• arguments:
  • --img_size
  • --conf_threshold
  • --iou_threshold

✔ Automatically performs

• Model loading (including .tar → .pth conversion when necessary)
• Preprocessing
• Inference
• Postprocessing & NMS
• Bounding-box drawing with class labels and scores on the output media

✔ Supports multiple input types

• Single image
• Folder of images
• Single video
• COCO-style dataset (test2017 + annotations)

✔ Saves all visualized outputs

• Images → pred_imgs/
• Videos → pred_videos/
• Datasets → pred_dataset/

Everything is saved in the user-defined output folder.

Describe alternatives you've considered

• Writing custom inference scripts for every experiment
• Testing in notebooks (not ideal for production or automation)
• Modifying internal repo modules before every run

These approaches work but are not scalable or user-friendly, and they lack a consistent visualization pipeline.
These methods work but are not efficient, scalable, or user-friendly.

Additional context

I have already written a working implementation of this inference script, which includes:

• CLI entry point
• Automatic path handling
• NMS
• Folder/video processing
• OpenCV visualization with bounding boxes, labels, and per-category counts
• .tar → .pth conversion for weights
• Unified prediction pipeline for all input types

This script could be added to the repo (e.g., in tools/ or root directory) to significantly streamline user workflow and improve accessibility.

If the maintainers approve, I can submit this as a PR.