Upload LLaVA-Next-3D/data_precessing/sam2_mask.py with huggingface_hub

LLaVA-Next-3D/data_precessing/sam2_mask.py

@@ -0,0 +1,100 @@
+from pycocotools import mask as maskUtils
+import numpy as np
+import matplotlib.pyplot as plt
+import glob
+import pdb
+import json
+import cv2
+import os
+from tqdm import tqdm
+from torch.utils.data import DataLoader, Dataset
+from decord import VideoReader
+from PIL import Image
+
+def decode_video_every_4_frame(video_path):
+    # Open the video file
+    vr = VideoReader(video_path, num_threads=1)
+    total_frames = len(vr)
+    frames = []
+    indices = [i for i in range(0, total_frames) if i % 4 == 0]
+    frames = vr.get_batch(indices).asnumpy()
+    frames = [Image.fromarray(frames[i]) for i in range(frames.shape[0])]
+    return frames
+
+def decode_video_at_fps(video_path, target_fps=4):
+    # Open the video file
+    cap = cv2.VideoCapture(video_path)
+    
+    if not cap.isOpened():
+        print("Error: Couldn't open video.")
+        return
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    frames = []
+    for frame_idx in range(total_frames):
+        ret, frame = cap.read()
+        if not ret:
+            break
+        if frame_idx % 4 == 0:
+            frames.append(frame)
+    cap.release()
+    return frames
+
+
+a = 0
+
+class SAVDataset(Dataset):
+    def __init__(self, data_paths):
+        self.data_paths = data_paths
+
+    def __len__(self):
+        return len(self.data_paths)
+
+    def __getitem__(self, idx):
+        path = self.data_paths[idx]
+        with open(path, 'r') as f:
+            item = json.load(f)
+            
+        masks = item['masklet']
+        video_id = item['video_id']
+    
+        output_folder = './extra_data/SA-V/sav_train/'+video_id[0:7]+'/'
+        # if os.path.exists(output_folder+ f'{video_id}_object_0.json'):
+        #     return 0
+
+        object_info = {}
+        for object_id in item['masklet_id']:
+            object_info[object_id] = {'size': [int(item['video_width']), int(item['video_height'])]}
+            
+        for idx, rle in enumerate(masks):
+            # rle = item['masklet'][idx]
+            binary_mask = maskUtils.decode(rle).transpose((2, 1, 0)) # [num, width, height]
+            for object_id, mask in enumerate(binary_mask):
+                if np.sum(mask) > 0:  # Check if the mask is not empty
+                    x1 = int(np.min(np.where(mask == 1)[0]))
+                    x2 = int(np.max(np.where(mask == 1)[0]))
+                    y1 = int(np.min(np.where(mask == 1)[1]))
+                    y2 = int(np.max(np.where(mask == 1)[1]))
+
+                    object_info[object_id].update({idx:[x1, y1, x2, y2]})
+
+        for object_id, info in object_info.items():
+            json.dump(info, open(output_folder+ f'{video_id}_object_{object_id}.json', 'w'), indent=4)
+
+        return 0
+
+all_paths = []
+data_paths = glob.glob('./extra_data/SA-V/sav_train/sav_*')
+for data_path in data_paths:
+    all_paths += glob.glob(data_path + '/*_manual.json')
+
+all_paths.sort()
+
+# Usage example:
+dataset = SAVDataset(all_paths)
+loader  = DataLoader(dataset, batch_size=16, shuffle=False, num_workers=16, drop_last=False)
+
+print(f"Number of dataset samples: {len(dataset)}")
+for idx, batch in enumerate(loader):
+    print(f'processing {idx}-th batch of {len(loader)}')
+    pass