Análise de video via GPT-40

.gitignore

@@ -4,3 +4,4 @@ video_analysis_output/
 get-pip.py
 *.m4a
 *.mp4
+audio_analysis_output/

requirements-video.txt

@@ -1 +1,2 @@
-yt-dlp opencv-python openai
+yt-dlp opencv-python openai
+torch torchvision transformers pillow scikit-learn

tool_video_analyzer.py

@@ -16,11 +16,16 @@ import shutil
 # --- Configurações (Substitua os placeholders) ---
 VIDEO_URL = "https://www.youtube.com/watch?v=L1vXCYZAYYM"  # Substitua pela URL do vídeo do YouTube
 OUTPUT_DIR = "./video_analysis_output" # Diretório para salvar o vídeo e os frames
-FRAME_INTERVAL_SECONDS = 3 # Intervalo entre frames a serem extraídos
+FRAME_INTERVAL_SECONDS = 1 # Intervalo entre frames a serem extraídos
+INICIO_FRAME_IMPORTANTE  = 96 # inicio intervalo relevante, para não ficar caro a inferencia ao gpt
+FIM_FRAME_IMPORTANTE = 96 # fim intervalo relevante, para não ficar caro a inferencia ao gpt
 OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
 GPT_MODEL = "gpt-4o" # Modelo GPT a ser usado (certifique-se que é o correto para análise de imagem)
 #PROMPT_TEXT = "You are an image analyzer, do not return any explanation. If asked to count items, return only an integer. If in doubt, return 0. How many different bird species are visible in the image?" # Prompt para o GPT-4o
-PROMPT_TEXT = "You are an expert in visual species classification. Based on the image provided, determine and return the number of distinct bird species visible. Do not count individuals — only count different species based on visual traits like size, shape, color, and beak structure. Return only a single integer. If unsure, return your best estimate. Do not provide explanations or any extra text."
+#PROMPT_TEXT = "You are an expert in visual species classification. Based on the image provided, determine and return the number of distinct bird species visible. Do not count individuals — only count different species based on visual traits like size, shape, color, and beak structure. Return only a single integer. If unsure, return your best estimate. Do not provide explanations or any extra text."
+#PROMPT_TEXT = "You are an expert in visual species classification. Based on the image provided, determine and return the number of distinct bird species visible. Do not count individuals — only count different species based on visual traits like size, shape, color, and beak structure. Return only a single integer and the species name. If unsure, return your best estimate. Do not provide explanations or any extra text."
+PROMPT_TEXT = "You are a world-class expert in avian species classification. Analyze the provided image and determine how many **distinct bird species** are present. Consider size, shape, plumage, coloration, and beak structure. Focus only on visible morphological differences. Return a **single integer** with no explanation. Do not count individuals of the same species. If unsure, return your most informed estimate."
+
 RESULTS_FILE = os.path.join(OUTPUT_DIR, "analysis_results.json")
 VIDEO_FILENAME = "downloaded_video.mp4"
 VIDEO_PATH = os.path.join(OUTPUT_DIR, VIDEO_FILENAME)
@@ -101,6 +106,11 @@ def extract_frames(video_path, output_dir, interval_sec):
         return []
 
     cap = cv2.VideoCapture(video_path)
+    # Verificar a resolução
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    print(f"Resolução original do vídeo: {width}x{height}")    
+    
     if not cap.isOpened():
         print(f"Erro ao abrir o arquivo de vídeo: {video_path}")
         return []
@@ -126,6 +136,10 @@ def extract_frames(video_path, output_dir, interval_sec):
         if target_frame_pos >= total_frames:
             break # Sai se o próximo frame alvo estiver além do final do vídeo
 
+        if (saved_frame_index < INICIO_FRAME_IMPORTANTE or saved_frame_index > FIM_FRAME_IMPORTANTE):
+            print(f"Pulando frame {saved_frame_index}")
+            saved_frame_index += 1
+            continue # evitar custo desnecessário para inferencia ao gpt
         cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame_pos)
         ret, frame = cap.read()
 
@@ -134,7 +148,8 @@ def extract_frames(video_path, output_dir, interval_sec):
             break # Sai se não conseguir ler o frame
 
         # redimensiona o frame (custo chamada)
-        frame = cv2.resize(frame, (1280, 720))
+        # removido porque poderia afetar a nitidez e impactar o resultado
+        # frame = cv2.resize(frame, (1280, 720))
         
         # Calcula o timestamp em segundos
         timestamp_sec = target_frame_pos / fps
@@ -143,7 +158,9 @@ def extract_frames(video_path, output_dir, interval_sec):
         frame_filename = f"frame_{saved_frame_index:04d}_time_{timestamp_sec:.2f}s.png"
         frame_path = os.path.join(output_dir, frame_filename)
         try:
-            cv2.imwrite(frame_path, frame)
+            # modificado para salvar com qualidade máxima cv2.imwrite(frame_path, frame)
+            cv2.imwrite(frame_path, frame, [cv2.IMWRITE_PNG_COMPRESSION, 0])
+
             extracted_frames_paths.append(frame_path)
             print(f"Frame salvo: {frame_path} (Timestamp: {timestamp_sec:.2f}s)")
             saved_frame_index += 1
@@ -292,7 +309,7 @@ if __name__ == "__main__":
         extracted_frames = extract_frames(VIDEO_PATH, OUTPUT_DIR, FRAME_INTERVAL_SECONDS)
     else:
         print("Pulando extração de frames pois o vídeo não está disponível.")
-
+    
     # Etapa 3 e 4: Codificar e Analisar Frames
     if extracted_frames and openai_client:
         print(f"\nIniciando análise de {len(extracted_frames)} frames com {GPT_MODEL}...")
@@ -310,7 +327,6 @@ if __name__ == "__main__":
 
             # Codifica o frame
             #teste com a imagem correta
-            frame_path = f"{OUTPUT_DIR}/frame_0031_time_93.00s.png"
             base64_image = encode_frame_to_base64(frame_path)
             
             if base64_image:
@@ -333,7 +349,7 @@ if __name__ == "__main__":
                     "analysis": {"error": "Failed to encode frame to base64."}
                 })
                 
-            break # teste somente uma chamada
+            # break # teste somente uma chamada
         print("\nAnálise de todos os frames concluída.")
     elif not extracted_frames:
         print("Nenhum frame foi extraído. Pulando etapa de análise.")