ultralytics 8.1.29__py3-none-any.whl → 8.3.62__py3-none-any.whl

ultralytics/data/base.py

@@ -1,4 +1,4 @@
-# Ultralytics YOLO 🚀, AGPL-3.0 license
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
 import glob
 import math
@@ -14,8 +14,8 @@ import numpy as np
 import psutil
 from torch.utils.data import Dataset
 
+from ultralytics.data.utils import FORMATS_HELP_MSG, HELP_URL, IMG_FORMATS
 from ultralytics.utils import DEFAULT_CFG, LOCAL_RANK, LOGGER, NUM_THREADS, TQDM
-from .utils import HELP_URL, IMG_FORMATS
 
 
 class BaseDataset(Dataset):
@@ -86,13 +86,19 @@ class BaseDataset(Dataset):
         self.buffer = []  # buffer size = batch size
         self.max_buffer_length = min((self.ni, self.batch_size * 8, 1000)) if self.augment else 0
 
-        # Cache images
-        if cache == "ram" and not self.check_cache_ram():
-            cache = False
+        # Cache images (options are cache = True, False, None, "ram", "disk")
         self.ims, self.im_hw0, self.im_hw = [None] * self.ni, [None] * self.ni, [None] * self.ni
         self.npy_files = [Path(f).with_suffix(".npy") for f in self.im_files]
-        if cache:
-            self.cache_images(cache)
+        self.cache = cache.lower() if isinstance(cache, str) else "ram" if cache is True else None
+        if self.cache == "ram" and self.check_cache_ram():
+            if hyp.deterministic:
+                LOGGER.warning(
+                    "WARNING ⚠️ cache='ram' may produce non-deterministic training results. "
+                    "Consider cache='disk' as a deterministic alternative if your disk space allows."
+                )
+            self.cache_images()
+        elif self.cache == "disk" and self.check_cache_disk():
+            self.cache_images()
 
         # Transforms
         self.transforms = self.build_transforms(hyp=hyp)
@@ -116,13 +122,11 @@ class BaseDataset(Dataset):
                     raise FileNotFoundError(f"{self.prefix}{p} does not exist")
             im_files = sorted(x.replace("/", os.sep) for x in f if x.split(".")[-1].lower() in IMG_FORMATS)
             # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in IMG_FORMATS])  # pathlib
-            assert im_files, f"{self.prefix}No images found in {img_path}"
+            assert im_files, f"{self.prefix}No images found in {img_path}. {FORMATS_HELP_MSG}"
         except Exception as e:
             raise FileNotFoundError(f"{self.prefix}Error loading data from {img_path}\n{HELP_URL}") from e
         if self.fraction < 1:
-            # im_files = im_files[: round(len(im_files) * self.fraction)]
-            num_elements_to_select = round(len(im_files) * self.fraction)
-            im_files = random.sample(im_files, num_elements_to_select)
+            im_files = im_files[: round(len(im_files) * self.fraction)]  # retain a fraction of the dataset
         return im_files
 
     def update_labels(self, include_class: Optional[list]):
@@ -173,28 +177,29 @@ class BaseDataset(Dataset):
             if self.augment:
                 self.ims[i], self.im_hw0[i], self.im_hw[i] = im, (h0, w0), im.shape[:2]  # im, hw_original, hw_resized
                 self.buffer.append(i)
-                if len(self.buffer) >= self.max_buffer_length:
+                if 1 < len(self.buffer) >= self.max_buffer_length:  # prevent empty buffer
                     j = self.buffer.pop(0)
-                    self.ims[j], self.im_hw0[j], self.im_hw[j] = None, None, None
+                    if self.cache != "ram":
+                        self.ims[j], self.im_hw0[j], self.im_hw[j] = None, None, None
 
             return im, (h0, w0), im.shape[:2]
 
         return self.ims[i], self.im_hw0[i], self.im_hw[i]
 
-    def cache_images(self, cache):
+    def cache_images(self):
         """Cache images to memory or disk."""
         b, gb = 0, 1 << 30  # bytes of cached images, bytes per gigabytes
-        fcn = self.cache_images_to_disk if cache == "disk" else self.load_image
+        fcn, storage = (self.cache_images_to_disk, "Disk") if self.cache == "disk" else (self.load_image, "RAM")
         with ThreadPool(NUM_THREADS) as pool:
             results = pool.imap(fcn, range(self.ni))
             pbar = TQDM(enumerate(results), total=self.ni, disable=LOCAL_RANK > 0)
             for i, x in pbar:
-                if cache == "disk":
+                if self.cache == "disk":
                     b += self.npy_files[i].stat().st_size
                 else:  # 'ram'
                     self.ims[i], self.im_hw0[i], self.im_hw[i] = x  # im, hw_orig, hw_resized = load_image(self, i)
                     b += self.ims[i].nbytes
-                pbar.desc = f"{self.prefix}Caching images ({b / gb:.1f}GB {cache})"
+                pbar.desc = f"{self.prefix}Caching images ({b / gb:.1f}GB {storage})"
             pbar.close()
 
     def cache_images_to_disk(self, i):
@@ -203,25 +208,55 @@ class BaseDataset(Dataset):
         if not f.exists():
             np.save(f.as_posix(), cv2.imread(self.im_files[i]), allow_pickle=False)
 
+    def check_cache_disk(self, safety_margin=0.5):
+        """Check image caching requirements vs available disk space."""
+        import shutil
+
+        b, gb = 0, 1 << 30  # bytes of cached images, bytes per gigabytes
+        n = min(self.ni, 30)  # extrapolate from 30 random images
+        for _ in range(n):
+            im_file = random.choice(self.im_files)
+            im = cv2.imread(im_file)
+            if im is None:
+                continue
+            b += im.nbytes
+            if not os.access(Path(im_file).parent, os.W_OK):
+                self.cache = None
+                LOGGER.info(f"{self.prefix}Skipping caching images to disk, directory not writeable ⚠️")
+                return False
+        disk_required = b * self.ni / n * (1 + safety_margin)  # bytes required to cache dataset to disk
+        total, used, free = shutil.disk_usage(Path(self.im_files[0]).parent)
+        if disk_required > free:
+            self.cache = None
+            LOGGER.info(
+                f"{self.prefix}{disk_required / gb:.1f}GB disk space required, "
+                f"with {int(safety_margin * 100)}% safety margin but only "
+                f"{free / gb:.1f}/{total / gb:.1f}GB free, not caching images to disk ⚠️"
+            )
+            return False
+        return True
+
     def check_cache_ram(self, safety_margin=0.5):
         """Check image caching requirements vs available memory."""
         b, gb = 0, 1 << 30  # bytes of cached images, bytes per gigabytes
         n = min(self.ni, 30)  # extrapolate from 30 random images
         for _ in range(n):
             im = cv2.imread(random.choice(self.im_files))  # sample image
+            if im is None:
+                continue
             ratio = self.imgsz / max(im.shape[0], im.shape[1])  # max(h, w)  # ratio
             b += im.nbytes * ratio**2
         mem_required = b * self.ni / n * (1 + safety_margin)  # GB required to cache dataset into RAM
         mem = psutil.virtual_memory()
-        cache = mem_required < mem.available  # to cache or not to cache, that is the question
-        if not cache:
+        if mem_required > mem.available:
+            self.cache = None
             LOGGER.info(
-                f'{self.prefix}{mem_required / gb:.1f}GB RAM required to cache images '
-                f'with {int(safety_margin * 100)}% safety margin but only '
-                f'{mem.available / gb:.1f}/{mem.total / gb:.1f}GB available, '
-                f"{'caching images ✅' if cache else 'not caching images ⚠️'}"
+                f"{self.prefix}{mem_required / gb:.1f}GB RAM required to cache images "
+                f"with {int(safety_margin * 100)}% safety margin but only "
+                f"{mem.available / gb:.1f}/{mem.total / gb:.1f}GB available, not caching images ⚠️"
             )
-        return cache
+            return False
+        return True
 
     def set_rectangle(self):
         """Sets the shape of bounding boxes for YOLO detections as rectangles."""
@@ -300,10 +335,10 @@ class BaseDataset(Dataset):
                 im_file=im_file,
                 shape=shape,  # format: (height, width)
                 cls=cls,
-                bboxes=bboxes, # xywh
+                bboxes=bboxes,  # xywh
                 segments=segments,  # xy
-                keypoints=keypoints, # xy
-                normalized=True, # or False
+                keypoints=keypoints,  # xy
+                normalized=True,  # or False
                 bbox_format="xyxy",  # or xywh, ltwh
             )
             ```

ultralytics/data/build.py

@@ -1,4 +1,4 @@
-# Ultralytics YOLO 🚀, AGPL-3.0 license
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
 import os
 import random
@@ -9,6 +9,7 @@ import torch
 from PIL import Image
 from torch.utils.data import dataloader, distributed
 
+from ultralytics.data.dataset import GroundingDataset, YOLODataset, YOLOMultiModalDataset
 from ultralytics.data.loaders import (
     LOADERS,
     LoadImagesAndVideos,
@@ -19,11 +20,9 @@ from ultralytics.data.loaders import (
     SourceTypes,
     autocast_list,
 )
-from ultralytics.data.utils import IMG_FORMATS, VID_FORMATS
+from ultralytics.data.utils import IMG_FORMATS, PIN_MEMORY, VID_FORMATS
 from ultralytics.utils import RANK, colorstr
 from ultralytics.utils.checks import check_file
-from .dataset import YOLODataset
-from .utils import PIN_MEMORY
 
 
 class InfiniteDataLoader(dataloader.DataLoader):
@@ -48,6 +47,13 @@ class InfiniteDataLoader(dataloader.DataLoader):
         for _ in range(len(self)):
             yield next(self.iterator)
 
+    def __del__(self):
+        """Ensure that workers are terminated."""
+        for w in self.iterator._workers:  # force terminate
+            if w.is_alive():
+                w.terminate()
+        self.iterator._shutdown_workers()  # cleanup
+
     def reset(self):
         """
         Reset iterator.
@@ -82,9 +88,10 @@ def seed_worker(worker_id):  # noqa
     random.seed(worker_seed)
 
 
-def build_yolo_dataset(cfg, img_path, batch, data, mode="train", rect=False, stride=32):
+def build_yolo_dataset(cfg, img_path, batch, data, mode="train", rect=False, stride=32, multi_modal=False):
     """Build YOLO Dataset."""
-    return YOLODataset(
+    dataset = YOLOMultiModalDataset if multi_modal else YOLODataset
+    return dataset(
         img_path=img_path,
         imgsz=cfg.imgsz,
         batch_size=batch,
@@ -103,11 +110,32 @@ def build_yolo_dataset(cfg, img_path, batch, data, mode="train", rect=False, str
     )
 
 
+def build_grounding(cfg, img_path, json_file, batch, mode="train", rect=False, stride=32):
+    """Build YOLO Dataset."""
+    return GroundingDataset(
+        img_path=img_path,
+        json_file=json_file,
+        imgsz=cfg.imgsz,
+        batch_size=batch,
+        augment=mode == "train",  # augmentation
+        hyp=cfg,  # TODO: probably add a get_hyps_from_cfg function
+        rect=cfg.rect or rect,  # rectangular batches
+        cache=cfg.cache or None,
+        single_cls=cfg.single_cls or False,
+        stride=int(stride),
+        pad=0.0 if mode == "train" else 0.5,
+        prefix=colorstr(f"{mode}: "),
+        task=cfg.task,
+        classes=cfg.classes,
+        fraction=cfg.fraction if mode == "train" else 1.0,
+    )
+
+
 def build_dataloader(dataset, batch, workers, shuffle=True, rank=-1):
     """Return an InfiniteDataLoader or DataLoader for training or validation set."""
     batch = min(batch, len(dataset))
     nd = torch.cuda.device_count()  # number of CUDA devices
-    nw = min([os.cpu_count() // max(nd, 1), workers])  # number of workers
+    nw = min(os.cpu_count() // max(nd, 1), workers)  # number of workers
     sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle)
     generator = torch.Generator()
     generator.manual_seed(6148914691236517205 + RANK)
@@ -129,7 +157,7 @@ def check_source(source):
     webcam, screenshot, from_img, in_memory, tensor = False, False, False, False, False
     if isinstance(source, (str, int, Path)):  # int for local usb camera
         source = str(source)
-        is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+        is_file = Path(source).suffix[1:] in (IMG_FORMATS | VID_FORMATS)
         is_url = source.lower().startswith(("https://", "http://", "rtsp://", "rtmp://", "tcp://"))
         webcam = source.isnumeric() or source.endswith(".streams") or (is_url and not is_file)
         screenshot = source.lower() == "screen"

ultralytics/data/converter.py

@@ -1,13 +1,18 @@
-# Ultralytics YOLO 🚀, AGPL-3.0 license
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
 import json
+import random
+import shutil
 from collections import defaultdict
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
 import cv2
 import numpy as np
+from PIL import Image
 
-from ultralytics.utils import LOGGER, TQDM
+from ultralytics.utils import DATASETS_DIR, LOGGER, NUM_THREADS, TQDM
+from ultralytics.utils.downloads import download
 from ultralytics.utils.files import increment_path
 
 
@@ -115,7 +120,7 @@ def coco91_to_coco80_class():
 
 
 def coco80_to_coco91_class():
-    """
+    r"""
     Converts 80-index (val2014) to 91-index (paper).
     For details see https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/.
 
@@ -123,8 +128,8 @@ def coco80_to_coco91_class():
         ```python
         import numpy as np
 
-        a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')
-        b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')
+        a = np.loadtxt("data/coco.names", dtype="str", delimiter="\n")
+        b = np.loadtxt("data/coco_paper.names", dtype="str", delimiter="\n")
         x1 = [list(a[i] == b).index(True) + 1 for i in range(80)]  # darknet to coco
         x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)]  # coco to darknet
         ```
@@ -219,6 +224,7 @@ def convert_coco(
     use_segments=False,
     use_keypoints=False,
     cls91to80=True,
+    lvis=False,
 ):
     """
     Converts COCO dataset annotations to a YOLO annotation format  suitable for training YOLO models.
@@ -229,18 +235,21 @@ def convert_coco(
         use_segments (bool, optional): Whether to include segmentation masks in the output.
         use_keypoints (bool, optional): Whether to include keypoint annotations in the output.
         cls91to80 (bool, optional): Whether to map 91 COCO class IDs to the corresponding 80 COCO class IDs.
+        lvis (bool, optional): Whether to convert data in lvis dataset way.
 
     Example:
         ```python
         from ultralytics.data.converter import convert_coco
 
-        convert_coco('../datasets/coco/annotations/', use_segments=True, use_keypoints=False, cls91to80=True)
+        convert_coco("../datasets/coco/annotations/", use_segments=True, use_keypoints=False, cls91to80=False)
+        convert_coco(
+            "../datasets/lvis/annotations/", use_segments=True, use_keypoints=False, cls91to80=False, lvis=True
+        )
         ```
 
     Output:
         Generates output files in the specified output directory.
     """
-
     # Create dataset directory
     save_dir = increment_path(save_dir)  # increment if save directory already exists
     for p in save_dir / "labels", save_dir / "images":
@@ -251,28 +260,38 @@ def convert_coco(
 
     # Import json
     for json_file in sorted(Path(labels_dir).resolve().glob("*.json")):
-        fn = Path(save_dir) / "labels" / json_file.stem.replace("instances_", "")  # folder name
+        lname = "" if lvis else json_file.stem.replace("instances_", "")
+        fn = Path(save_dir) / "labels" / lname  # folder name
         fn.mkdir(parents=True, exist_ok=True)
-        with open(json_file) as f:
+        if lvis:
+            # NOTE: create folders for both train and val in advance,
+            # since LVIS val set contains images from COCO 2017 train in addition to the COCO 2017 val split.
+            (fn / "train2017").mkdir(parents=True, exist_ok=True)
+            (fn / "val2017").mkdir(parents=True, exist_ok=True)
+        with open(json_file, encoding="utf-8") as f:
             data = json.load(f)
 
         # Create image dict
-        images = {f'{x["id"]:d}': x for x in data["images"]}
+        images = {f"{x['id']:d}": x for x in data["images"]}
         # Create image-annotations dict
         imgToAnns = defaultdict(list)
         for ann in data["annotations"]:
             imgToAnns[ann["image_id"]].append(ann)
 
+        image_txt = []
         # Write labels file
         for img_id, anns in TQDM(imgToAnns.items(), desc=f"Annotations {json_file}"):
             img = images[f"{img_id:d}"]
-            h, w, f = img["height"], img["width"], img["file_name"]
+            h, w = img["height"], img["width"]
+            f = str(Path(img["coco_url"]).relative_to("http://images.cocodataset.org")) if lvis else img["file_name"]
+            if lvis:
+                image_txt.append(str(Path("./images") / f))
 
             bboxes = []
             segments = []
             keypoints = []
             for ann in anns:
-                if ann["iscrowd"]:
+                if ann.get("iscrowd", False):
                     continue
                 # The COCO box format is [top left x, top left y, width, height]
                 box = np.array(ann["bbox"], dtype=np.float64)
@@ -314,7 +333,89 @@ def convert_coco(
                         )  # cls, box or segments
                     file.write(("%g " * len(line)).rstrip() % line + "\n")
 
-    LOGGER.info(f"COCO data converted successfully.\nResults saved to {save_dir.resolve()}")
+        if lvis:
+            with open((Path(save_dir) / json_file.name.replace("lvis_v1_", "").replace(".json", ".txt")), "a") as f:
+                f.writelines(f"{line}\n" for line in image_txt)
+
+    LOGGER.info(f"{'LVIS' if lvis else 'COCO'} data converted successfully.\nResults saved to {save_dir.resolve()}")
+
+
+def convert_segment_masks_to_yolo_seg(masks_dir, output_dir, classes):
+    """
+    Converts a dataset of segmentation mask images to the YOLO segmentation format.
+
+    This function takes the directory containing the binary format mask images and converts them into YOLO segmentation format.
+    The converted masks are saved in the specified output directory.
+
+    Args:
+        masks_dir (str): The path to the directory where all mask images (png, jpg) are stored.
+        output_dir (str): The path to the directory where the converted YOLO segmentation masks will be stored.
+        classes (int): Total classes in the dataset i.e. for COCO classes=80
+
+    Example:
+        ```python
+        from ultralytics.data.converter import convert_segment_masks_to_yolo_seg
+
+        # The classes here is the total classes in the dataset, for COCO dataset we have 80 classes
+        convert_segment_masks_to_yolo_seg("path/to/masks_directory", "path/to/output/directory", classes=80)
+        ```
+
+    Notes:
+        The expected directory structure for the masks is:
+
+            - masks
+                ├─ mask_image_01.png or mask_image_01.jpg
+                ├─ mask_image_02.png or mask_image_02.jpg
+                ├─ mask_image_03.png or mask_image_03.jpg
+                └─ mask_image_04.png or mask_image_04.jpg
+
+        After execution, the labels will be organized in the following structure:
+
+            - output_dir
+                ├─ mask_yolo_01.txt
+                ├─ mask_yolo_02.txt
+                ├─ mask_yolo_03.txt
+                └─ mask_yolo_04.txt
+    """
+    pixel_to_class_mapping = {i + 1: i for i in range(classes)}
+    for mask_path in Path(masks_dir).iterdir():
+        if mask_path.suffix in {".png", ".jpg"}:
+            mask = cv2.imread(str(mask_path), cv2.IMREAD_GRAYSCALE)  # Read the mask image in grayscale
+            img_height, img_width = mask.shape  # Get image dimensions
+            LOGGER.info(f"Processing {mask_path} imgsz = {img_height} x {img_width}")
+
+            unique_values = np.unique(mask)  # Get unique pixel values representing different classes
+            yolo_format_data = []
+
+            for value in unique_values:
+                if value == 0:
+                    continue  # Skip background
+                class_index = pixel_to_class_mapping.get(value, -1)
+                if class_index == -1:
+                    LOGGER.warning(f"Unknown class for pixel value {value} in file {mask_path}, skipping.")
+                    continue
+
+                # Create a binary mask for the current class and find contours
+                contours, _ = cv2.findContours(
+                    (mask == value).astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
+                )  # Find contours
+
+                for contour in contours:
+                    if len(contour) >= 3:  # YOLO requires at least 3 points for a valid segmentation
+                        contour = contour.squeeze()  # Remove single-dimensional entries
+                        yolo_format = [class_index]
+                        for point in contour:
+                            # Normalize the coordinates
+                            yolo_format.append(round(point[0] / img_width, 6))  # Rounding to 6 decimal places
+                            yolo_format.append(round(point[1] / img_height, 6))
+                        yolo_format_data.append(yolo_format)
+            # Save Ultralytics YOLO format data to file
+            output_path = Path(output_dir) / f"{mask_path.stem}.txt"
+            with open(output_path, "w") as file:
+                for item in yolo_format_data:
+                    line = " ".join(map(str, item))
+                    file.write(line + "\n")
+            LOGGER.info(f"Processed and stored at {output_path} imgsz = {img_height} x {img_width}")
 
 
 def convert_dota_to_yolo_obb(dota_root_path: str):
@@ -331,7 +432,7 @@ def convert_dota_to_yolo_obb(dota_root_path: str):
         ```python
         from ultralytics.data.converter import convert_dota_to_yolo_obb
 
-        convert_dota_to_yolo_obb('path/to/DOTA')
+        convert_dota_to_yolo_obb("path/to/DOTA")
         ```
 
     Notes:
@@ -393,7 +494,7 @@ def convert_dota_to_yolo_obb(dota_root_path: str):
                 normalized_coords = [
                     coords[i] / image_width if i % 2 == 0 else coords[i] / image_height for i in range(8)
                 ]
-                formatted_coords = ["{:.6g}".format(coord) for coord in normalized_coords]
+                formatted_coords = [f"{coord:.6g}" for coord in normalized_coords]
                 g.write(f"{class_idx} {' '.join(formatted_coords)}\n")
 
     for phase in ["train", "val"]:
@@ -463,7 +564,7 @@ def merge_multi_segment(segments):
                 segments[i] = np.roll(segments[i], -idx[0], axis=0)
                 segments[i] = np.concatenate([segments[i], segments[i][:1]])
                 # Deal with the first segment and the last one
-                if i in [0, len(idx_list) - 1]:
+                if i in {0, len(idx_list) - 1}:
                     s.append(segments[i])
                 else:
                     idx = [0, idx[1] - idx[0]]
@@ -471,14 +572,14 @@ def merge_multi_segment(segments):
 
         else:
             for i in range(len(idx_list) - 1, -1, -1):
-                if i not in [0, len(idx_list) - 1]:
+                if i not in {0, len(idx_list) - 1}:
                     idx = idx_list[i]
                     nidx = abs(idx[1] - idx[0])
                     s.append(segments[i][nidx:])
     return s
 
 
-def yolo_bbox2segment(im_dir, save_dir=None, sam_model="sam_b.pt"):
+def yolo_bbox2segment(im_dir, save_dir=None, sam_model="sam_b.pt", device=None):
     """
     Converts existing object detection dataset (bounding boxes) to segmentation dataset or oriented bounding box (OBB)
     in YOLO format. Generates segmentation data using SAM auto-annotator as needed.
@@ -488,24 +589,24 @@ def yolo_bbox2segment(im_dir, save_dir=None, sam_model="sam_b.pt"):
         save_dir (str | Path): Path to save the generated labels, labels will be saved
             into `labels-segment` in the same directory level of `im_dir` if save_dir is None. Default: None.
         sam_model (str): Segmentation model to use for intermediate segmentation data; optional.
+        device (int | str): The specific device to run SAM models. Default: None.
 
     Notes:
         The input directory structure assumed for dataset:
 
             - im_dir
                 ├─ 001.jpg
-                ├─ ..
+                ├─ ...
                 └─ NNN.jpg
             - labels
                 ├─ 001.txt
-                ├─ ..
+                ├─ ...
                 └─ NNN.txt
     """
+    from ultralytics import SAM
     from ultralytics.data import YOLODataset
-    from ultralytics.utils.ops import xywh2xyxy
     from ultralytics.utils import LOGGER
-    from ultralytics import SAM
-    from tqdm import tqdm
+    from ultralytics.utils.ops import xywh2xyxy
 
     # NOTE: add placeholder to pass class index check
     dataset = YOLODataset(im_dir, data=dict(names=list(range(1000))))
@@ -515,28 +616,87 @@ def yolo_bbox2segment(im_dir, save_dir=None, sam_model="sam_b.pt"):
 
     LOGGER.info("Detection labels detected, generating segment labels by SAM model!")
     sam_model = SAM(sam_model)
-    for l in tqdm(dataset.labels, total=len(dataset.labels), desc="Generating segment labels"):
-        h, w = l["shape"]
-        boxes = l["bboxes"]
+    for label in TQDM(dataset.labels, total=len(dataset.labels), desc="Generating segment labels"):
+        h, w = label["shape"]
+        boxes = label["bboxes"]
         if len(boxes) == 0:  # skip empty labels
             continue
         boxes[:, [0, 2]] *= w
         boxes[:, [1, 3]] *= h
-        im = cv2.imread(l["im_file"])
-        sam_results = sam_model(im, bboxes=xywh2xyxy(boxes), verbose=False, save=False)
-        l["segments"] = sam_results[0].masks.xyn
+        im = cv2.imread(label["im_file"])
+        sam_results = sam_model(im, bboxes=xywh2xyxy(boxes), verbose=False, save=False, device=device)
+        label["segments"] = sam_results[0].masks.xyn
 
     save_dir = Path(save_dir) if save_dir else Path(im_dir).parent / "labels-segment"
     save_dir.mkdir(parents=True, exist_ok=True)
-    for l in dataset.labels:
+    for label in dataset.labels:
         texts = []
-        lb_name = Path(l["im_file"]).with_suffix(".txt").name
+        lb_name = Path(label["im_file"]).with_suffix(".txt").name
         txt_file = save_dir / lb_name
-        cls = l["cls"]
-        for i, s in enumerate(l["segments"]):
+        cls = label["cls"]
+        for i, s in enumerate(label["segments"]):
+            if len(s) == 0:
+                continue
             line = (int(cls[i]), *s.reshape(-1))
             texts.append(("%g " * len(line)).rstrip() % line)
-        if texts:
-            with open(txt_file, "a") as f:
-                f.writelines(text + "\n" for text in texts)
+        with open(txt_file, "a") as f:
+            f.writelines(text + "\n" for text in texts)
     LOGGER.info(f"Generated segment labels saved in {save_dir}")
+
+
+def create_synthetic_coco_dataset():
+    """
+    Creates a synthetic COCO dataset with random images based on filenames from label lists.
+
+    This function downloads COCO labels, reads image filenames from label list files,
+    creates synthetic images for train2017 and val2017 subsets, and organizes
+    them in the COCO dataset structure. It uses multithreading to generate images efficiently.
+
+    Examples:
+        >>> from ultralytics.data.converter import create_synthetic_coco_dataset
+        >>> create_synthetic_coco_dataset()
+
+    Notes:
+        - Requires internet connection to download label files.
+        - Generates random RGB images of varying sizes (480x480 to 640x640 pixels).
+        - Existing test2017 directory is removed as it's not needed.
+        - Reads image filenames from train2017.txt and val2017.txt files.
+    """
+
+    def create_synthetic_image(image_file):
+        """Generates synthetic images with random sizes and colors for dataset augmentation or testing purposes."""
+        if not image_file.exists():
+            size = (random.randint(480, 640), random.randint(480, 640))
+            Image.new(
+                "RGB",
+                size=size,
+                color=(random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)),
+            ).save(image_file)
+
+    # Download labels
+    dir = DATASETS_DIR / "coco"
+    url = "https://github.com/ultralytics/assets/releases/download/v0.0.0/"
+    label_zip = "coco2017labels-segments.zip"
+    download([url + label_zip], dir=dir.parent)
+
+    # Create synthetic images
+    shutil.rmtree(dir / "labels" / "test2017", ignore_errors=True)  # Remove test2017 directory as not needed
+    with ThreadPoolExecutor(max_workers=NUM_THREADS) as executor:
+        for subset in ["train2017", "val2017"]:
+            subset_dir = dir / "images" / subset
+            subset_dir.mkdir(parents=True, exist_ok=True)
+
+            # Read image filenames from label list file
+            label_list_file = dir / f"{subset}.txt"
+            if label_list_file.exists():
+                with open(label_list_file) as f:
+                    image_files = [dir / line.strip() for line in f]
+
+                # Submit all tasks
+                futures = [executor.submit(create_synthetic_image, image_file) for image_file in image_files]
+                for _ in TQDM(as_completed(futures), total=len(futures), desc=f"Generating images for {subset}"):
+                    pass  # The actual work is done in the background
+            else:
+                print(f"Warning: Labels file {label_list_file} does not exist. Skipping image creation for {subset}.")
+
+    print("Synthetic COCO dataset created successfully.")