Update app.py

app.py

@@ -1,24 +1,77 @@
 import gradio as gr
 from transformers import pipeline
+import cv2  # OpenCV for video processing
 
-# Replace with a suitable image classification model ID
-model_id = "sayakpaul/resnet-50-finetuned-imagenet"
+# Model ID for video classification (UCF101 subset)
+model_id = "sayakpaul/videomae-base-finetuned-ucf101-subset"
 
-def analyze_image(image):
-    classifier = pipeline("image-classification", model=model_id)
-    predictions = classifier(images=image)  # Assuming the model outputs probabilities
-    # Extract the most likely class and its probability
-    top_class = predictions[0]["label"]
-    top_prob = predictions[0]["score"]
-    return f"Top Class: {top_class} (Probability: {top_prob:.2f})"
+def analyze_video(video):
+    # Extract key frames from the video using OpenCV
+    frames = extract_key_frames(video)
+    
+    # Analyze key frames using video classification model
+    results = []
+    classifier = pipeline("video-classification", model=model_id)
+    for frame in frames:
+        predictions = classifier(images=frame)  # Assuming model outputs probabilities
+        # Analyze predictions for insights related to the play
+        result = analyze_predictions_ucf101(predictions)
+        results.append(result)
+
+    # Aggregate results across frames and provide a final analysis
+    final_result = aggregate_results(results)
+
+    return final_result
+
+def extract_key_frames(video):
+    cap = cv2.VideoCapture(video)
+    frames = []
+    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = int(cap.get(cv2.CAP_PROP_FPS))
+    
+    for i in range(frame_count):
+        ret, frame = cap.read()
+        if ret and i % (fps // 2) == 0:  # Extract a frame every half second
+            frames.append(frame)
+    
+    cap.release()
+    return frames
+
+def analyze_predictions_ucf101(predictions):
+    # Analyze the model's predictions (probabilities) for insights relevant to baseball plays
+    # For simplicity, we'll assume predictions return the top-1 class
+    actions = [pred['label'] for pred in predictions]
+    
+    relevant_actions = ["running", "sliding", "jumping"]
+    runner_actions = [action for action in actions if action in relevant_actions]
+    
+    # Check for 'running', 'sliding' actions as key indicators for safe/out decision
+    if "sliding" in runner_actions:
+        return "potentially safe"
+    elif "running" in runner_actions:
+        return "potentially out"
+    else:
+        return "inconclusive"
+
+def aggregate_results(results):
+    # Combine insights from analyzing each frame (e.g., dominant action classes, confidence scores)
+    safe_count = results.count("potentially safe")
+    out_count = results.count("potentially out")
+    
+    if safe_count > out_count:
+        return "Safe"
+    elif out_count > safe_count:
+        return "Out"
+    else:
+        return "Inconclusive"
 
 # Gradio interface
 interface = gr.Interface(
-    fn=analyze_image,
-    inputs="image",
+    fn=analyze_video,
+    inputs="video",
     outputs="text",
-    title="Image Analyzer (Generic)",
-    description="Upload an image and get the most likely classification based on the chosen model.",
+    title="Baseball Play Analysis (UCF101 Subset Exploration)",
+    description="Upload a video of a baseball play (safe/out at a base). This app explores using a video classification model (UCF101 subset) for analysis. Note: The model might not be specifically trained for baseball plays.",
 )
 
 interface.launch()