momentry_core/scripts/zero_shot_gun_test.py

#!/opt/homebrew/bin/python3.11
"""
Zero-shot Gun Detection Test — OWL-ViT vs Grounding DINO
Tests on 8 known timepoints: 5 original pistol frames + 3 ASR gun mentions.
"""
import json, os, sys, time, cv2
import torch
from PIL import Image
import numpy as np

VIDEO = "/Users/accusys/momentry/var/sftpgo/data/demo/Charade (1963) Cary Grant & Audrey Hepburn \uff5c Comedy Mystery Romance Thriller \uff5c Full Movie.mp4"
OUTPUT_DIR = "/Users/accusys/momentry/output_dev/zero_shot_test"
os.makedirs(OUTPUT_DIR, exist_ok=True)

TIMEPOINTS = [
    (2646, "2646s", "ASR: He has a gun"),
    (3188, "3188s", "Original pistol"),
    (3697, "3697s", "ASR: Where's your gun"),
    (5341, "5341s", "ASR: He already killed 3 men"),
    (5461, "5461s", "Original pistol"),
    (6309, "6309s", "Original pistol"),
    (6377, "6377s", "Original gun"),
    (6479, "6479s", "Original pistol"),
]
PROMPTS = ["gun", "pistol", "rifle", "weapon"]

cap = cv2.VideoCapture(VIDEO)
fps = cap.get(cv2.CAP_PROP_FPS) or 25.0

def get_frame(t_sec):
    cap.set(cv2.CAP_PROP_POS_FRAMES, int(t_sec * fps))
    ret, frame = cap.read()
    return frame if ret else None

def save_annotated(frame, detections, prompt, model_name, label):
    img = frame.copy()
    for d in detections:
        x1, y1, x2, y2 = [int(v) for v in d["bbox"]]
        conf = d["score"]
        cls = d["label"]
        cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv2.putText(img, f"{cls} {conf:.2f}", (x1, y1-5),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
    filename = f"{label}_{model_name}_prompt-{prompt}.jpg"
    cv2.imwrite(os.path.join(OUTPUT_DIR, filename), img, [cv2.IMWRITE_JPEG_QUALITY, 85])
    return filename

all_results = {}

# ========== OWL-ViT ==========
print("=" * 60)
print("OWL-ViT (google/owlvit-base-patch32)")
print("=" * 60)

from transformers import OwlViTProcessor, OwlViTForObjectDetection

owl_proc = OwlViTProcessor.from_pretrained("google/owlvit-base-patch32")
owl_model = OwlViTForObjectDetection.from_pretrained("google/owlvit-base-patch32")
device = "mps" if torch.backends.mps.is_available() else "cpu"
owl_model.to(device)
print(f"Device: {device}")

owl_dets = {}
t0 = time.time()
for t_sec, label, desc in TIMEPOINTS:
    frame = get_frame(t_sec)
    if frame is None: continue
    img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    for prompt in PROMPTS:
        inputs = owl_proc(text=[[prompt]], images=img, return_tensors="pt").to(device)
        with torch.no_grad():
            outputs = owl_model(**inputs)
        target = torch.tensor([img.size[::-1]])
        dets = owl_proc.post_process_grounded_object_detection(outputs, threshold=0.05, target_sizes=target)[0]
        det_list = []
        for i in range(len(dets["boxes"])):
            det_list.append({
                "bbox": [round(v, 1) for v in dets["boxes"][i].tolist()],
                "score": round(dets["scores"][i].item(), 3),
                "label": prompt,
            })
        save_annotated(frame, det_list, prompt, "owlvit", label)
        key = f"{label}_prompt-{prompt}"
        owl_dets[key] = det_list
        if det_list:
            best = max(d["score"] for d in det_list)
            print(f"  [{desc}] prompt='{prompt}': {len(det_list)} det best={best:.3f}")

all_results["owlvit"] = {"elapsed": round(time.time()-t0, 1), "detections": owl_dets}

# ========== Grounding DINO ==========
print("\n" + "=" * 60)
print("Grounding DINO (IDEA-Research/grounding-dino-base)")
print("=" * 60)

from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection

gd_proc = AutoProcessor.from_pretrained("IDEA-Research/grounding-dino-base")
gd_model = AutoModelForZeroShotObjectDetection.from_pretrained("IDEA-Research/grounding-dino-base")
gd_model.to(device)

gd_dets = {}
t0 = time.time()
for t_sec, label, desc in TIMEPOINTS:
    frame = get_frame(t_sec)
    if frame is None: continue
    img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    for prompt in PROMPTS:
        inputs = gd_proc(images=img, text=prompt, return_tensors="pt").to(device)
        with torch.no_grad():
            outputs = gd_model(**inputs)
        target = torch.tensor([img.size[::-1]])
        dets = gd_proc.post_process_grounded_object_detection(outputs, threshold=0.05, target_sizes=target)[0]
        det_list = []
        for i in range(len(dets["boxes"])):
            det_list.append({
                "bbox": [round(v, 1) for v in dets["boxes"][i].tolist()],
                "score": round(dets["scores"][i].item(), 3),
                "label": prompt,
            })
        save_annotated(frame, det_list, prompt, "grounding-dino", label)
        key = f"{label}_prompt-{prompt}"
        gd_dets[key] = det_list
        if det_list:
            best = max(d["score"] for d in det_list)
            print(f"  [{desc}] prompt='{prompt}': {len(det_list)} det best={best:.3f}")

all_results["grounding-dino"] = {"elapsed": round(time.time()-t0, 1), "detections": gd_dets}

cap.release()

# ========== Summary ==========
print("\n" + "=" * 60)
print("SUMMARY")
print("=" * 60)
for model in ["owlvit", "grounding-dino"]:
    d = all_results[model]
    dets = d["detections"]
    hits = sum(1 for v in dets.values() if v)
    total = sum(len(v) for v in dets.values())
    print(f"\n{model} ({d['elapsed']}s): {hits}/8 timepoints, {total} total detections")
    for t_sec, label, desc in TIMEPOINTS:
        candidates = []
        for p in PROMPTS:
            key = f"{label}_prompt-{p}"
            if key in dets and dets[key]:
                for dd in dets[key]:
                    candidates.append((p, dd["score"]))
        if candidates:
            best = max(candidates, key=lambda x: x[1])
            print(f"  {desc}: best={best[1]:.3f} (prompt='{best[0]}')")
        else:
            print(f"  {desc}: no detections")

json.dump(all_results, open(os.path.join(OUTPUT_DIR, "zero_shot_results.json"), "w"), indent=2)
print(f"\nSaved to {OUTPUT_DIR}/")