detect_objects.py

import argparse
import json
import numpy as np
import torch


def json_default(obj):
    if isinstance(obj, (np.integer,)):
        return int(obj)
    if isinstance(obj, (np.floating,)):
        return float(obj)
    if isinstance(obj, np.ndarray):
        return obj.tolist()
    raise TypeError(f"Object of type {type(obj).__name__} is not JSON serializable")

from sam3.model_builder import build_sam3_video_predictor


def main():
    parser = argparse.ArgumentParser(description="Detect objects in video first frame using SAM 3 text prompt")
    parser.add_argument("--video_path", required=True, help="Path to video (MP4 or JPEG folder)")
    parser.add_argument("--text", required=True, help="Text prompt for detection")
    parser.add_argument("--frame_index", type=int, default=0, help="Frame index to prompt on")
    parser.add_argument("--output_json", required=True, help="Output JSON path")
    args = parser.parse_args()

    # Build model (downloads from HF automatically if not cached)
    video_predictor = build_sam3_video_predictor()

    # Start session
    response = video_predictor.handle_request(
        request=dict(
            type="start_session",
            resource_path=args.video_path,
        )
    )
    session_id = response["session_id"]

    # Add text prompt on the specified frame
    response = video_predictor.handle_request(
        request=dict(
            type="add_prompt",
            session_id=session_id,
            frame_index=args.frame_index,
            text=args.text,
        )
    )

    outputs = response["outputs"]
    obj_ids = outputs["out_obj_ids"]
    probs = outputs["out_probs"]
    boxes_xywh = outputs["out_boxes_xywh"]
    binary_masks = outputs["out_binary_masks"]

    # Collect results
    instances = []
    for i in range(len(obj_ids)):
        obj_id = obj_ids[i]
        if torch.is_tensor(obj_id):
            obj_id = int(obj_id.detach().cpu())

        box = boxes_xywh[i]
        if torch.is_tensor(box):
            box = box.detach().cpu().tolist()

        # Convert xywh to xyxy
        x, y, w, h = box
        bbox_xyxy = [x, y, x + w, y + h]

        prob = probs[i]
        if torch.is_tensor(prob):
            prob = float(prob.detach().cpu().max()) if prob.numel() > 1 else float(prob.detach().cpu())

        instances.append(dict(id=obj_id, bbox_xyxy=bbox_xyxy, bbox_xywh=box, score=prob))

    result = dict(
        video_path=args.video_path,
        text=args.text,
        frame_index=args.frame_index,
        num_instances=len(instances),
        instances=instances,
    )

    with open(args.output_json, "w") as f:
        json.dump(result, f, indent=2, default=json_default)

    print(f"Found {len(instances)} objects for prompt '{args.text}'")
    print(f"Saved to {args.output_json}")


if __name__ == "__main__":
    main()