Update sam2

segment-anything-2/sam2/sam2_video_predictor.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the root directory of this source tree.
 
+import warnings
 from collections import OrderedDict
 
 import torch
@@ -44,11 +45,13 @@ class SAM2VideoPredictor(SAM2Base):
         async_loading_frames=False,
     ):
         """Initialize a inference state."""
+        compute_device = self.device  # device of the model
         images, video_height, video_width = load_video_frames(
             video_path=video_path,
             image_size=self.image_size,
             offload_video_to_cpu=offload_video_to_cpu,
             async_loading_frames=async_loading_frames,
+            compute_device=compute_device,
         )
         inference_state = {}
         inference_state["images"] = images
@@ -64,11 +67,11 @@ class SAM2VideoPredictor(SAM2Base):
         # the original video height and width, used for resizing final output scores
         inference_state["video_height"] = video_height
         inference_state["video_width"] = video_width
-        inference_state["device"] = torch.device("cuda")
+        inference_state["device"] = compute_device
         if offload_state_to_cpu:
             inference_state["storage_device"] = torch.device("cpu")
         else:
-            inference_state["storage_device"] = torch.device("cuda")
+            inference_state["storage_device"] = compute_device
         # inputs on each frame
         inference_state["point_inputs_per_obj"] = {}
         inference_state["mask_inputs_per_obj"] = {}
@@ -103,6 +106,23 @@ class SAM2VideoPredictor(SAM2Base):
         self._get_image_feature(inference_state, frame_idx=0, batch_size=1)
         return inference_state
 
+    @classmethod
+    def from_pretrained(cls, model_id: str, **kwargs) -> "SAM2VideoPredictor":
+        """
+        Load a pretrained model from the Hugging Face hub.
+
+        Arguments:
+          model_id (str): The Hugging Face repository ID.
+          **kwargs: Additional arguments to pass to the model constructor.
+
+        Returns:
+          (SAM2VideoPredictor): The loaded model.
+        """
+        from sam2.build_sam import build_sam2_video_predictor_hf
+
+        sam_model = build_sam2_video_predictor_hf(model_id, **kwargs)
+        return sam_model
+
     def _obj_id_to_idx(self, inference_state, obj_id):
         """Map client-side object id to model-side object index."""
         obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
@@ -146,29 +166,66 @@ class SAM2VideoPredictor(SAM2Base):
         return len(inference_state["obj_idx_to_id"])
 
     @torch.inference_mode()
-    def add_new_points(
+    def add_new_points_or_box(
         self,
         inference_state,
         frame_idx,
         obj_id,
-        points,
-        labels,
+        points=None,
+        labels=None,
         clear_old_points=True,
         normalize_coords=True,
+        box=None,
     ):
         """Add new points to a frame."""
         obj_idx = self._obj_id_to_idx(inference_state, obj_id)
         point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
         mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
 
-        if not isinstance(points, torch.Tensor):
+        if (points is not None) != (labels is not None):
+            raise ValueError("points and labels must be provided together")
+        if points is None and box is None:
+            raise ValueError("at least one of points or box must be provided as input")
+
+        if points is None:
+            points = torch.zeros(0, 2, dtype=torch.float32)
+        elif not isinstance(points, torch.Tensor):
             points = torch.tensor(points, dtype=torch.float32)
-        if not isinstance(labels, torch.Tensor):
+        if labels is None:
+            labels = torch.zeros(0, dtype=torch.int32)
+        elif not isinstance(labels, torch.Tensor):
             labels = torch.tensor(labels, dtype=torch.int32)
         if points.dim() == 2:
             points = points.unsqueeze(0)  # add batch dimension
         if labels.dim() == 1:
             labels = labels.unsqueeze(0)  # add batch dimension
+
+        # If `box` is provided, we add it as the first two points with labels 2 and 3
+        # along with the user-provided points (consistent with how SAM 2 is trained).
+        if box is not None:
+            if not clear_old_points:
+                raise ValueError(
+                    "cannot add box without clearing old points, since "
+                    "box prompt must be provided before any point prompt "
+                    "(please use clear_old_points=True instead)"
+                )
+            if inference_state["tracking_has_started"]:
+                warnings.warn(
+                    "You are adding a box after tracking starts. SAM 2 may not always be "
+                    "able to incorporate a box prompt for *refinement*. If you intend to "
+                    "use box prompt as an *initial* input before tracking, please call "
+                    "'reset_state' on the inference state to restart from scratch.",
+                    category=UserWarning,
+                    stacklevel=2,
+                )
+            if not isinstance(box, torch.Tensor):
+                box = torch.tensor(box, dtype=torch.float32, device=points.device)
+            box_coords = box.reshape(1, 2, 2)
+            box_labels = torch.tensor([2, 3], dtype=torch.int32, device=labels.device)
+            box_labels = box_labels.reshape(1, 2)
+            points = torch.cat([box_coords, points], dim=1)
+            labels = torch.cat([box_labels, labels], dim=1)
+
         if normalize_coords:
             video_H = inference_state["video_height"]
             video_W = inference_state["video_width"]
@@ -215,7 +272,8 @@ class SAM2VideoPredictor(SAM2Base):
                 prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
 
         if prev_out is not None and prev_out["pred_masks"] is not None:
-            prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
+            device = inference_state["device"]
+            prev_sam_mask_logits = prev_out["pred_masks"].to(device, non_blocking=True)
             # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
             prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
         current_out, _ = self._run_single_frame_inference(
@@ -251,6 +309,10 @@ class SAM2VideoPredictor(SAM2Base):
         )
         return frame_idx, obj_ids, video_res_masks
 
+    def add_new_points(self, *args, **kwargs):
+        """Deprecated method. Please use `add_new_points_or_box` instead."""
+        return self.add_new_points_or_box(*args, **kwargs)
+
     @torch.inference_mode()
     def add_new_mask(
         self,
@@ -531,7 +593,7 @@ class SAM2VideoPredictor(SAM2Base):
             storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
             # Find all the frames that contain temporary outputs for any objects
             # (these should be the frames that have just received clicks for mask inputs
-            # via `add_new_points` or `add_new_mask`)
+            # via `add_new_points_or_box` or `add_new_mask`)
             temp_frame_inds = set()
             for obj_temp_output_dict in temp_output_dict_per_obj.values():
                 temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
@@ -734,7 +796,8 @@ class SAM2VideoPredictor(SAM2Base):
         )
         if backbone_out is None:
             # Cache miss -- we will run inference on a single image
-            image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
+            device = inference_state["device"]
+            image = inference_state["images"][frame_idx].to(device).float().unsqueeze(0)
             backbone_out = self.forward_image(image)
             # Cache the most recent frame's feature (for repeated interactions with
             # a frame; we can use an LRU cache for more frames in the future).
@@ -895,4 +958,4 @@ class SAM2VideoPredictor(SAM2Base):
         for t in range(frame_idx_begin, frame_idx_end + 1):
             non_cond_frame_outputs.pop(t, None)
             for obj_output_dict in inference_state["output_dict_per_obj"].values():
-                obj_output_dict["non_cond_frame_outputs"].pop(t, None)
+                obj_output_dict["non_cond_frame_outputs"].pop(t, None)

segment-anything-2/sam2/utils/misc.py

@@ -106,7 +106,15 @@ class AsyncVideoFrameLoader:
     A list of video frames to be load asynchronously without blocking session start.
     """
 
-    def __init__(self, img_paths, image_size, offload_video_to_cpu, img_mean, img_std):
+    def __init__(
+        self,
+        img_paths,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        compute_device,
+    ):
         self.img_paths = img_paths
         self.image_size = image_size
         self.offload_video_to_cpu = offload_video_to_cpu
@@ -119,6 +127,7 @@ class AsyncVideoFrameLoader:
         # video_height and video_width be filled when loading the first image
         self.video_height = None
         self.video_width = None
+        self.compute_device = compute_device
 
         # load the first frame to fill video_height and video_width and also
         # to cache it (since it's most likely where the user will click)
@@ -152,7 +161,7 @@ class AsyncVideoFrameLoader:
         img -= self.img_mean
         img /= self.img_std
         if not self.offload_video_to_cpu:
-            img = img.cuda(non_blocking=True)
+            img = img.to(self.compute_device, non_blocking=True)
         self.images[index] = img
         return img
 
@@ -167,6 +176,7 @@ def load_video_frames(
     img_mean=(0.485, 0.456, 0.406),
     img_std=(0.229, 0.224, 0.225),
     async_loading_frames=False,
+    compute_device=torch.device("cuda"),
 ):
     """
     Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).
@@ -179,7 +189,15 @@ def load_video_frames(
     if isinstance(video_path, str) and os.path.isdir(video_path):
         jpg_folder = video_path
     else:
-        raise NotImplementedError("Only JPEG frames are supported at this moment")
+        raise NotImplementedError(
+            "Only JPEG frames are supported at this moment. For video files, you may use "
+            "ffmpeg (https://ffmpeg.org/) to extract frames into a folder of JPEG files, such as \n"
+            "```\n"
+            "ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <output_dir>/'%05d.jpg'\n"
+            "```\n"
+            "where `-q:v` generates high-quality JPEG frames and `-start_number 0` asks "
+            "ffmpeg to start the JPEG file from 00000.jpg."
+        )
 
     frame_names = [
         p
@@ -196,7 +214,12 @@ def load_video_frames(
 
     if async_loading_frames:
         lazy_images = AsyncVideoFrameLoader(
-            img_paths, image_size, offload_video_to_cpu, img_mean, img_std
+            img_paths,
+            image_size,
+            offload_video_to_cpu,
+            img_mean,
+            img_std,
+            compute_device,
         )
         return lazy_images, lazy_images.video_height, lazy_images.video_width
 
@@ -204,9 +227,9 @@ def load_video_frames(
     for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
         images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
     if not offload_video_to_cpu:
-        images = images.cuda()
-        img_mean = img_mean.cuda()
-        img_std = img_std.cuda()
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
     # normalize by mean and std
     images -= img_mean
     images /= img_std
@@ -220,10 +243,25 @@ def fill_holes_in_mask_scores(mask, max_area):
     # Holes are those connected components in background with area <= self.max_area
     # (background regions are those with mask scores <= 0)
     assert max_area > 0, "max_area must be positive"
-    labels, areas = get_connected_components(mask <= 0)
-    is_hole = (labels > 0) & (areas <= max_area)
-    # We fill holes with a small positive mask score (0.1) to change them to foreground.
-    mask = torch.where(is_hole, 0.1, mask)
+
+    input_mask = mask
+    try:
+        labels, areas = get_connected_components(mask <= 0)
+        is_hole = (labels > 0) & (areas <= max_area)
+        # We fill holes with a small positive mask score (0.1) to change them to foreground.
+        mask = torch.where(is_hole, 0.1, mask)
+    except Exception as e:
+        # Skip the post-processing step on removing small holes if the CUDA kernel fails
+        warnings.warn(
+            f"{e}\n\nSkipping the post-processing step due to the error above. You can "
+            "still use SAM 2 and it's OK to ignore the error above, although some post-processing "
+            "functionality may be limited (which doesn't affect the results in most cases; see "
+            "https://github.com/facebookresearch/segment-anything-2/blob/main/INSTALL.md).",
+            category=UserWarning,
+            stacklevel=2,
+        )
+        mask = input_mask
+
     return mask
 
 
@@ -235,4 +273,4 @@ def concat_points(old_point_inputs, new_points, new_labels):
         points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
         labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
 
-    return {"point_coords": points, "point_labels": labels}
+    return {"point_coords": points, "point_labels": labels}