Merge branch 'main' into yolo-serve

.github/workflows/ci.yaml

@@ -39,7 +39,8 @@ on:
 
 jobs:
   HUB:
-    if: github.repository == 'ultralytics/ultralytics' && (github.event_name == 'schedule' || github.event_name == 'push' || (github.event_name == 'workflow_dispatch' && github.event.inputs.hub == 'true'))
+    # if: github.repository == 'ultralytics/ultralytics' && (github.event_name == 'schedule' || github.event_name == 'push' || (github.event_name == 'workflow_dispatch' && github.event.inputs.hub == 'true'))
+    if: github.repository == 'ultralytics/ultralytics' && 'workflow_dispatch' && github.event.inputs.hub == 'true'
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false

.github/workflows/docker.yaml

@@ -84,11 +84,8 @@ jobs:
     outputs:
       new_release: ${{ steps.check_tag.outputs.new_release }}
     steps:
-      - name: Cleanup disk
+      - name: Cleanup disk space
-        # Free up to 30GB of disk space per https://github.com/ultralytics/ultralytics/pull/15848
+        uses: ultralytics/actions/cleanup-disk@main
-        uses: jlumbroso/free-disk-space@v1.3.1
-        with:
-          tool-cache: true
 
       - name: Checkout repo
         uses: actions/checkout@v4

README.md

@@ -17,6 +17,7 @@
     <a href="https://console.paperspace.com/github/ultralytics/ultralytics"><img src="https://assets.paperspace.io/img/gradient-badge.svg" alt="Run Ultralytics on Gradient"></a>
     <a href="https://colab.research.google.com/github/ultralytics/ultralytics/blob/main/examples/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open Ultralytics In Colab"></a>
     <a href="https://www.kaggle.com/models/ultralytics/yolo11"><img src="https://kaggle.com/static/images/open-in-kaggle.svg" alt="Open Ultralytics In Kaggle"></a>
+    <a href="https://mybinder.org/v2/gh/ultralytics/ultralytics/HEAD?labpath=examples%2Ftutorial.ipynb"><img src="https://mybinder.org/badge_logo.svg" alt="Open Ultralytics In Binder"></a>
 </div>
 <br>
 

README.zh-CN.md

@@ -17,6 +17,7 @@
     <a href="https://console.paperspace.com/github/ultralytics/ultralytics"><img src="https://assets.paperspace.io/img/gradient-badge.svg" alt="Run Ultralytics on Gradient"></a>
     <a href="https://colab.research.google.com/github/ultralytics/ultralytics/blob/main/examples/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open Ultralytics In Colab"></a>
     <a href="https://www.kaggle.com/models/ultralytics/yolo11"><img src="https://kaggle.com/static/images/open-in-kaggle.svg" alt="Open Ultralytics In Kaggle"></a>
+    <a href="https://mybinder.org/v2/gh/ultralytics/ultralytics/HEAD?labpath=examples%2Ftutorial.ipynb"><img src="https://mybinder.org/badge_logo.svg" alt="Open Ultralytics In Binder"></a>
 </div>
 <br>
 

docs/en/datasets/index.md

@@ -19,7 +19,11 @@ Ultralytics provides support for various datasets to facilitate computer vision
   <strong>Watch:</strong> Ultralytics Datasets Overview
 </p>
 
-## Ultralytics Explorer 🚀 NEW
+## Ultralytics Explorer
+
+!!! warning "Community Note ⚠️"
+
+    As of **`ultralytics>=8.3.10`**, Ultralytics explorer support has been deprecated. But don't worry! You can now access similar and even enhanced functionality through [Ultralytics HUB](https://hub.ultralytics.com/), our intuitive no-code platform designed to streamline your workflow. With Ultralytics HUB, you can continue exploring, visualizing, and managing your data effortlessly, all without writing a single line of code. Make sure to check it out and take advantage of its powerful features!🚀
 
 Create [embeddings](https://www.ultralytics.com/glossary/embeddings) for your dataset, search for similar images, run SQL queries, perform semantic search and even search using natural language! You can get started with our GUI app or build your own using the API. Learn more [here](explorer/index.md).
 

docs/en/models/sam-2.md

@@ -1,9 +1,13 @@
 ---
 comments: true
 description: Discover SAM 2, the next generation of Meta's Segment Anything Model, supporting real-time promptable segmentation in both images and videos with state-of-the-art performance. Learn about its key features, datasets, and how to use it.
-keywords: SAM 2, Segment Anything, video segmentation, image segmentation, promptable segmentation, zero-shot performance, SA-V dataset, Ultralytics, real-time segmentation, AI, machine learning
+keywords: SAM 2, SAM 2.1, Segment Anything, video segmentation, image segmentation, promptable segmentation, zero-shot performance, SA-V dataset, Ultralytics, real-time segmentation, AI, machine learning
 ---
 
+!!! tip "SAM 2.1"
+
+    We have just supported the more accurate SAM2.1 model. Please give it a try!
+
 # SAM 2: Segment Anything Model 2
 
 SAM 2, the successor to Meta's [Segment Anything Model (SAM)](sam.md), is a cutting-edge tool designed for comprehensive object segmentation in both images and videos. It excels in handling complex visual data through a unified, promptable model architecture that supports real-time processing and zero-shot generalization.
@@ -114,12 +118,16 @@ pip install ultralytics
 
 The following table details the available SAM 2 models, their pre-trained weights, supported tasks, and compatibility with different operating modes like [Inference](../modes/predict.md), [Validation](../modes/val.md), [Training](../modes/train.md), and [Export](../modes/export.md).
 
-| Model Type  | Pre-trained Weights                                                                   | Tasks Supported                              | Inference | Validation | Training | Export |
+| Model Type    | Pre-trained Weights                                                                       | Tasks Supported                              | Inference | Validation | Training | Export |
-| ----------- | ------------------------------------------------------------------------------------- | -------------------------------------------- | --------- | ---------- | -------- | ------ |
+| ------------- | ----------------------------------------------------------------------------------------- | -------------------------------------------- | --------- | ---------- | -------- | ------ |
-| SAM 2 tiny  | [sam2_t.pt](https://github.com/ultralytics/assets/releases/download/v8.2.0/sam2_t.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2 tiny    | [sam2_t.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2_t.pt)     | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
-| SAM 2 small | [sam2_s.pt](https://github.com/ultralytics/assets/releases/download/v8.2.0/sam2_s.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2 small   | [sam2_s.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2_s.pt)     | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
-| SAM 2 base  | [sam2_b.pt](https://github.com/ultralytics/assets/releases/download/v8.2.0/sam2_b.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2 base    | [sam2_b.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2_b.pt)     | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
-| SAM 2 large | [sam2_l.pt](https://github.com/ultralytics/assets/releases/download/v8.2.0/sam2_l.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2 large   | [sam2_l.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2_l.pt)     | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2.1 tiny  | [sam2.1_t.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2.1_t.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2.1 small | [sam2.1_s.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2.1_s.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2.1 base  | [sam2.1_b.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2.1_b.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
+| SAM 2.1 large | [sam2.1_l.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/sam2.1_l.pt) | [Instance Segmentation](../tasks/segment.md) | ✅        | ❌         | ❌       | ❌     |
 
 ### SAM 2 Prediction Examples
 
@@ -137,7 +145,7 @@ SAM 2 can be utilized across a broad spectrum of tasks, including real-time vide
         from ultralytics import SAM
 
         # Load a model
-        model = SAM("sam2_b.pt")
+        model = SAM("sam2.1_b.pt")
 
         # Display model information (optional)
         model.info()
@@ -170,7 +178,7 @@ SAM 2 can be utilized across a broad spectrum of tasks, including real-time vide
         from ultralytics import SAM
 
         # Load a model
-        model = SAM("sam2_b.pt")
+        model = SAM("sam2.1_b.pt")
 
         # Display model information (optional)
         model.info()
@@ -183,7 +191,7 @@ SAM 2 can be utilized across a broad spectrum of tasks, including real-time vide
 
         ```bash
         # Run inference with a SAM 2 model
-        yolo predict model=sam2_b.pt source=path/to/video.mp4
+        yolo predict model=sam2.1_b.pt source=path/to/video.mp4
         ```
 
 - This example demonstrates how SAM 2 can be used to segment the entire content of an image or video if no prompts (bboxes/points/masks) are provided.

docs/en/modes/export.md

@@ -136,13 +136,13 @@ INT8 quantization is an excellent way to compress the model and speed up inferen
         from ultralytics import YOLO
 
         model = YOLO("yolo11n.pt")  # Load a model
-        model.export(format="onnx", int8=True)
+        model.export(format="engine", int8=True)
         ```
 
     === "CLI"
 
         ```bash
-        yolo export model=yolo11n.pt format=onnx int8=True   # export model with INT8 quantization
+        yolo export model=yolo11n.pt format=engine int8=True   # export TensorRT model with INT8 quantization
         ```
 
 INT8 quantization can be applied to various formats, such as TensorRT and CoreML. More details can be found in the [Export section](../modes/export.md).

docs/en/modes/predict.md

@@ -665,7 +665,7 @@ For more details see the [`Probs` class documentation](../reference/engine/resul
     model = YOLO("yolo11n-obb.pt")
 
     # Run inference on an image
-    results = model("bus.jpg")  # results list
+    results = model("boats.jpg")  # results list
 
     # View results
     for r in results:

docs/en/tasks/obb.md

@@ -141,14 +141,14 @@ Use a trained YOLO11n-obb model to run predictions on images.
         model = YOLO("path/to/best.pt")  # load a custom model
 
         # Predict with the model
-        results = model("https://ultralytics.com/images/bus.jpg")  # predict on an image
+        results = model("https://ultralytics.com/images/boats.jpg")  # predict on an image
         ```
 
     === "CLI"
 
         ```bash
-        yolo obb predict model=yolo11n-obb.pt source='https://ultralytics.com/images/bus.jpg'  # predict with official model
+        yolo obb predict model=yolo11n-obb.pt source='https://ultralytics.com/images/boats.jpg'  # predict with official model
-        yolo obb predict model=path/to/best.pt source='https://ultralytics.com/images/bus.jpg'  # predict with custom model
+        yolo obb predict model=path/to/best.pt source='https://ultralytics.com/images/boats.jpg'  # predict with custom model
         ```
 
 <p align="center">

examples/tutorial.ipynb

@@ -583,7 +583,7 @@
         "\n",
         "model = YOLO('yolo11n-obb.pt')  # load a pretrained YOLO OBB model\n",
         "model.train(data='dota8.yaml', epochs=3)  # train the model\n",
-        "model('https://ultralytics.com/images/bus.jpg')  # predict on an image"
+        "model('https://ultralytics.com/images/boats.jpg')  # predict on an image"
       ],
       "metadata": {
         "id": "IJNKClOOB5YS"

pyproject.toml

@@ -61,7 +61,8 @@ classifiers = [
 
 # Required dependencies ------------------------------------------------------------------------------------------------
 dependencies = [
-    "numpy>=1.23.0", # temporary patch for compat errors https://github.com/ultralytics/yolov5/actions/runs/9538130424/job/26286956354
+    "numpy>=1.23.0",
+    "numpy<2.0.0; sys_platform == 'darwin'", # macOS OpenVINO errors https://github.com/ultralytics/ultralytics/pull/17221
     "matplotlib>=3.3.0",
     "opencv-python>=4.6.0",
     "pillow>=7.1.2",

tests/test_cuda.py

@@ -116,7 +116,7 @@ def test_predict_sam():
     from ultralytics.models.sam import Predictor as SAMPredictor
 
     # Load a model
-    model = SAM(WEIGHTS_DIR / "sam_b.pt")
+    model = SAM(WEIGHTS_DIR / "sam2.1_b.pt")
 
     # Display model information (optional)
     model.info()

ultralytics/__init__.py

@@ -1,6 +1,6 @@
 # Ultralytics YOLO 🚀, AGPL-3.0 license
 
-__version__ = "8.3.21"
+__version__ = "8.3.23"
 
 import os
 

ultralytics/cfg/__init__.py

@@ -802,7 +802,7 @@ def entrypoint(debug=""):
         from ultralytics import FastSAM
 
         model = FastSAM(model)
-    elif "sam_" in stem or "sam2_" in stem:
+    elif "sam_" in stem or "sam2_" in stem or "sam2.1_" in stem:
         from ultralytics import SAM
 
         model = SAM(model)
@@ -824,7 +824,9 @@ def entrypoint(debug=""):
 
     # Mode
     if mode in {"predict", "track"} and "source" not in overrides:
-        overrides["source"] = DEFAULT_CFG.source or ASSETS
+        overrides["source"] = (
+            "https://ultralytics.com/images/boats.jpg" if task == "obb" else DEFAULT_CFG.source or ASSETS
+        )
         LOGGER.warning(f"WARNING ⚠️ 'source' argument is missing. Using default 'source={overrides['source']}'.")
     elif mode in {"train", "val"}:
         if "data" not in overrides and "resume" not in overrides:

ultralytics/cfg/solutions/default.yaml

@@ -1,18 +1,19 @@
 # Ultralytics YOLO 🚀, AGPL-3.0 license
-
 # Configuration for Ultralytics Solutions
 
-model: "yolo11n.pt" # The Ultralytics YOLO11 model to be used (e.g., yolo11n.pt for YOLO11 nano version and yolov8n.pt for YOLOv8 nano version)
+# Object counting settings
-
 region: # Object counting, queue or speed estimation region points. Default region points are [(20, 400), (1080, 404), (1080, 360), (20, 360)]
-line_width: 2 # Width of the annotator used to draw regions on the image/video frames + bounding boxes and tracks drawing. Default value is 2.
-show: True # Flag to control whether to display output image or not, you can set this as False i.e. when deploying it on some embedded devices.
 show_in: True # Flag to display objects moving *into* the defined region
 show_out: True # Flag to display objects moving *out of* the defined region
-classes: # To count specific classes. i.e, if you want to detect, track and count the person with COCO model, you can use classes=0, Default its None
+
+# Heatmaps settings
+colormap: # Colormap for heatmap, Only OPENCV supported colormaps can be used. By default COLORMAP_PARULA will be used for visualization.
+
+# Workouts monitoring settings
 up_angle: 145.0 # Workouts up_angle for counts, 145.0 is default value. You can adjust it for different workouts, based on position of keypoints.
 down_angle: 90 # Workouts down_angle for counts, 90 is default value. You can change it for different workouts, based on position of keypoints.
 kpts: [6, 8, 10] # Keypoints for workouts monitoring, i.e. If you want to consider keypoints for pushups that have mostly values of [6, 8, 10].
-colormap: # Colormap for heatmap, Only OPENCV supported colormaps can be used. By default COLORMAP_PARULA will be used for visualization.
+
+# Analytics settings
 analytics_type: "line" # Analytics type i.e "line", "pie", "bar" or "area" charts. By default, "line" analytics will be used for processing.
 json_file: # parking system regions file path.

ultralytics/data/converter.py

@@ -632,9 +632,10 @@ def yolo_bbox2segment(im_dir, save_dir=None, sam_model="sam_b.pt"):
         txt_file = save_dir / lb_name
         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)
     LOGGER.info(f"Generated segment labels saved in {save_dir}")

ultralytics/engine/exporter.py

@@ -457,6 +457,7 @@ class Exporter:
     @try_export
     def export_openvino(self, prefix=colorstr("OpenVINO:")):
         """YOLO OpenVINO export."""
+        # WARNING: numpy>=2.0.0 issue with OpenVINO on macOS https://github.com/ultralytics/ultralytics/pull/17221
         check_requirements(f'openvino{"<=2024.0.0" if ARM64 else ">=2024.0.0"}')  # fix OpenVINO issue on ARM64
         import openvino as ov
 

ultralytics/models/sam/build.py

@@ -263,6 +263,7 @@ def _build_sam2(
     memory_attention = MemoryAttention(d_model=256, pos_enc_at_input=True, num_layers=4, layer=MemoryAttentionLayer())
     memory_encoder = MemoryEncoder(out_dim=64)
 
+    is_sam2_1 = checkpoint is not None and "sam2.1" in checkpoint
     sam2 = SAM2Model(
         image_encoder=image_encoder,
         memory_attention=memory_attention,
@@ -288,6 +289,9 @@ def _build_sam2(
         multimask_max_pt_num=1,
         use_mlp_for_obj_ptr_proj=True,
         compile_image_encoder=False,
+        no_obj_embed_spatial=is_sam2_1,
+        proj_tpos_enc_in_obj_ptrs=is_sam2_1,
+        use_signed_tpos_enc_to_obj_ptrs=is_sam2_1,
         sam_mask_decoder_extra_args=dict(
             dynamic_multimask_via_stability=True,
             dynamic_multimask_stability_delta=0.05,
@@ -313,6 +317,10 @@ sam_model_map = {
     "sam2_s.pt": build_sam2_s,
     "sam2_b.pt": build_sam2_b,
     "sam2_l.pt": build_sam2_l,
+    "sam2.1_t.pt": build_sam2_t,
+    "sam2.1_s.pt": build_sam2_s,
+    "sam2.1_b.pt": build_sam2_b,
+    "sam2.1_l.pt": build_sam2_l,
 }
 
 

ultralytics/models/sam/modules/sam.py

@@ -161,18 +161,19 @@ class SAM2Model(torch.nn.Module):
         use_multimask_token_for_obj_ptr: bool = False,
         iou_prediction_use_sigmoid=False,
         memory_temporal_stride_for_eval=1,
-        add_all_frames_to_correct_as_cond=False,
         non_overlap_masks_for_mem_enc=False,
         use_obj_ptrs_in_encoder=False,
         max_obj_ptrs_in_encoder=16,
         add_tpos_enc_to_obj_ptrs=True,
         proj_tpos_enc_in_obj_ptrs=False,
+        use_signed_tpos_enc_to_obj_ptrs=False,
         only_obj_ptrs_in_the_past_for_eval=False,
         pred_obj_scores: bool = False,
         pred_obj_scores_mlp: bool = False,
         fixed_no_obj_ptr: bool = False,
         soft_no_obj_ptr: bool = False,
         use_mlp_for_obj_ptr_proj: bool = False,
+        no_obj_embed_spatial: bool = False,
         sam_mask_decoder_extra_args=None,
         compile_image_encoder: bool = False,
     ):
@@ -205,8 +206,6 @@ class SAM2Model(torch.nn.Module):
             use_multimask_token_for_obj_ptr (bool): Whether to use multimask tokens for object pointers.
             iou_prediction_use_sigmoid (bool): Whether to use sigmoid to restrict IoU prediction to [0-1].
             memory_temporal_stride_for_eval (int): Memory bank's temporal stride during evaluation.
-            add_all_frames_to_correct_as_cond (bool): Whether to append frames with correction clicks to conditioning
-                frame list.
             non_overlap_masks_for_mem_enc (bool): Whether to apply non-overlapping constraints on object masks in
                 memory encoder during evaluation.
             use_obj_ptrs_in_encoder (bool): Whether to cross-attend to object pointers from other frames in the encoder.
@@ -216,6 +215,9 @@ class SAM2Model(torch.nn.Module):
                 the encoder.
             proj_tpos_enc_in_obj_ptrs (bool): Whether to add an extra linear projection layer for temporal positional
                 encoding in object pointers.
+            use_signed_tpos_enc_to_obj_ptrs (bool): whether to use signed distance (instead of unsigned absolute distance)
+                in the temporal positional encoding in the object pointers, only relevant when both `use_obj_ptrs_in_encoder=True`
+                and `add_tpos_enc_to_obj_ptrs=True`.
             only_obj_ptrs_in_the_past_for_eval (bool): Whether to only attend to object pointers in the past
                 during evaluation.
             pred_obj_scores (bool): Whether to predict if there is an object in the frame.
@@ -223,6 +225,7 @@ class SAM2Model(torch.nn.Module):
             fixed_no_obj_ptr (bool): Whether to have a fixed no-object pointer when there is no object present.
             soft_no_obj_ptr (bool): Whether to mix in no-object pointer softly for easier recovery and error mitigation.
             use_mlp_for_obj_ptr_proj (bool): Whether to use MLP for object pointer projection.
+            no_obj_embed_spatial (bool): Whether add no obj embedding to spatial frames.
             sam_mask_decoder_extra_args (Dict | None): Extra arguments for constructing the SAM mask decoder.
             compile_image_encoder (bool): Whether to compile the image encoder for faster inference.
 
@@ -253,6 +256,7 @@ class SAM2Model(torch.nn.Module):
         if proj_tpos_enc_in_obj_ptrs:
             assert add_tpos_enc_to_obj_ptrs  # these options need to be used together
         self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
+        self.use_signed_tpos_enc_to_obj_ptrs = use_signed_tpos_enc_to_obj_ptrs
         self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval
 
         # Part 2: memory attention to condition current frame's visual features
@@ -309,9 +313,12 @@ class SAM2Model(torch.nn.Module):
             self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
             trunc_normal_(self.no_obj_ptr, std=0.02)
         self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj
+        self.no_obj_embed_spatial = None
+        if no_obj_embed_spatial:
+            self.no_obj_embed_spatial = torch.nn.Parameter(torch.zeros(1, self.mem_dim))
+            trunc_normal_(self.no_obj_embed_spatial, std=0.02)
 
         self._build_sam_heads()
-        self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
         self.max_cond_frames_in_attn = max_cond_frames_in_attn
 
         # Model compilation
@@ -533,8 +540,6 @@ class SAM2Model(torch.nn.Module):
         if self.pred_obj_scores:
             # Allow *soft* no obj ptr, unlike for masks
             if self.soft_no_obj_ptr:
-                # Only hard possible with gt
-                assert not self.teacher_force_obj_scores_for_mem
                 lambda_is_obj_appearing = object_score_logits.sigmoid()
             else:
                 lambda_is_obj_appearing = is_obj_appearing.float()
@@ -647,6 +652,7 @@ class SAM2Model(torch.nn.Module):
         if self.num_maskmem == 0:  # Disable memory and skip fusion
             return current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
         num_obj_ptr_tokens = 0
+        tpos_sign_mul = -1 if track_in_reverse else 1
         # Step 1: condition the visual features of the current frame on previous memories
         if not is_init_cond_frame:
             # Retrieve the memories encoded with the maskmem backbone
@@ -664,7 +670,7 @@ class SAM2Model(torch.nn.Module):
             # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
             # We also allow taking the memory frame non-consecutively (with r>1), in which case
             # we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
-            r = self.memory_temporal_stride_for_eval
+            r = 1 if self.training else self.memory_temporal_stride_for_eval
             for t_pos in range(1, self.num_maskmem):
                 t_rel = self.num_maskmem - t_pos  # how many frames before current frame
                 if t_rel == 1:
@@ -718,7 +724,14 @@ class SAM2Model(torch.nn.Module):
                     ptr_cond_outputs = selected_cond_outputs
                 pos_and_ptrs = [
                     # Temporal pos encoding contains how far away each pointer is from current frame
-                    (abs(frame_idx - t), out["obj_ptr"])
+                    (
+                        (
+                            (frame_idx - t) * tpos_sign_mul
+                            if self.use_signed_tpos_enc_to_obj_ptrs
+                            else abs(frame_idx - t)
+                        ),
+                        out["obj_ptr"],
+                    )
                     for t, out in ptr_cond_outputs.items()
                 ]
                 # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
@@ -787,6 +800,7 @@ class SAM2Model(torch.nn.Module):
         current_vision_feats,
         feat_sizes,
         pred_masks_high_res,
+        object_score_logits,
         is_mask_from_pts,
     ):
         """Encodes frame features and masks into a new memory representation for video segmentation."""
@@ -819,10 +833,17 @@ class SAM2Model(torch.nn.Module):
         )
         maskmem_features = maskmem_out["vision_features"]
         maskmem_pos_enc = maskmem_out["vision_pos_enc"]
+        # add a no-object embedding to the spatial memory to indicate that the frame
+        # is predicted to be occluded (i.e. no object is appearing in the frame)
+        if self.no_obj_embed_spatial is not None:
+            is_obj_appearing = (object_score_logits > 0).float()
+            maskmem_features += (1 - is_obj_appearing[..., None, None]) * self.no_obj_embed_spatial[
+                ..., None, None
+            ].expand(*maskmem_features.shape)
 
         return maskmem_features, maskmem_pos_enc
 
-    def track_step(
+    def _track_step(
         self,
         frame_idx,
         is_init_cond_frame,
@@ -833,15 +854,7 @@ class SAM2Model(torch.nn.Module):
         mask_inputs,
         output_dict,
         num_frames,
-        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        prev_sam_mask_logits,
-        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
-        # to skip the memory encoder with `run_mem_encoder=False`. For example,
-        # in demo we might call `track_step` multiple times for each user click,
-        # and only encode the memory when the user finalizes their clicks. And in ablation
-        # settings like SAM training on static images, we don't need the memory encoder.
-        run_mem_encoder=True,
-        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
-        prev_sam_mask_logits=None,
     ):
         """Performs a single tracking step, updating object masks and memory features based on current frame inputs."""
         current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
@@ -861,7 +874,7 @@ class SAM2Model(torch.nn.Module):
             sam_outputs = self._use_mask_as_output(pix_feat, high_res_features, mask_inputs)
         else:
             # fused the visual feature with previous memory features in the memory bank
-            pix_feat_with_mem = self._prepare_memory_conditioned_features(
+            pix_feat = self._prepare_memory_conditioned_features(
                 frame_idx=frame_idx,
                 is_init_cond_frame=is_init_cond_frame,
                 current_vision_feats=current_vision_feats[-1:],
@@ -880,12 +893,78 @@ class SAM2Model(torch.nn.Module):
                 mask_inputs = prev_sam_mask_logits
             multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
             sam_outputs = self._forward_sam_heads(
-                backbone_features=pix_feat_with_mem,
+                backbone_features=pix_feat,
                 point_inputs=point_inputs,
                 mask_inputs=mask_inputs,
                 high_res_features=high_res_features,
                 multimask_output=multimask_output,
             )
+        return current_out, sam_outputs, high_res_features, pix_feat
+
+    def _encode_memory_in_output(
+        self,
+        current_vision_feats,
+        feat_sizes,
+        point_inputs,
+        run_mem_encoder,
+        high_res_masks,
+        object_score_logits,
+        current_out,
+    ):
+        """Finally run the memory encoder on the predicted mask to encode, it into a new memory feature (that can be
+        used in future frames).
+        """
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                object_score_logits=object_score_logits,
+                is_mask_from_pts=(point_inputs is not None),
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+    ):
+        """Performs a single tracking step, updating object masks and memory features based on current frame inputs."""
+        current_out, sam_outputs, _, _ = self._track_step(
+            frame_idx,
+            is_init_cond_frame,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+            point_inputs,
+            mask_inputs,
+            output_dict,
+            num_frames,
+            track_in_reverse,
+            prev_sam_mask_logits,
+        )
+
         (
             _,
             _,
@@ -893,28 +972,28 @@ class SAM2Model(torch.nn.Module):
             low_res_masks,
             high_res_masks,
             obj_ptr,
-            _,
+            object_score_logits,
         ) = sam_outputs
 
         current_out["pred_masks"] = low_res_masks
         current_out["pred_masks_high_res"] = high_res_masks
         current_out["obj_ptr"] = obj_ptr
+        if not self.training:
+            # Only add this in inference (to avoid unused param in activation checkpointing;
+            # it's mainly used in the demo to encode spatial memories w/ consolidated masks)
+            current_out["object_score_logits"] = object_score_logits
 
         # Finally run the memory encoder on the predicted mask to encode
         # it into a new memory feature (that can be used in future frames)
-        if run_mem_encoder and self.num_maskmem > 0:
+        self._encode_memory_in_output(
-            high_res_masks_for_mem_enc = high_res_masks
+            current_vision_feats,
-            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            feat_sizes,
-                current_vision_feats=current_vision_feats,
+            point_inputs,
-                feat_sizes=feat_sizes,
+            run_mem_encoder,
-                pred_masks_high_res=high_res_masks_for_mem_enc,
+            high_res_masks,
-                is_mask_from_pts=(point_inputs is not None),
+            object_score_logits,
-            )
+            current_out,
-            current_out["maskmem_features"] = maskmem_features
+        )
-            current_out["maskmem_pos_enc"] = maskmem_pos_enc
-        else:
-            current_out["maskmem_features"] = None
-            current_out["maskmem_pos_enc"] = None
 
         return current_out
 

ultralytics/solutions/parking_management.py

@@ -168,7 +168,6 @@ class ParkingManagement(BaseSolution):
     Examples:
         >>> from ultralytics.solutions import ParkingManagement
         >>> parking_manager = ParkingManagement(model="yolov8n.pt", json_file="parking_regions.json")
-        >>> results = parking_manager(source="parking_lot_video.mp4")
         >>> print(f"Occupied spaces: {parking_manager.pr_info['Occupancy']}")
         >>> print(f"Available spaces: {parking_manager.pr_info['Available']}")
     """

ultralytics/solutions/solutions.py

@@ -1,16 +1,13 @@
 # Ultralytics YOLO 🚀, AGPL-3.0 license
 
 from collections import defaultdict
-from pathlib import Path
 
 import cv2
 
 from ultralytics import YOLO
-from ultralytics.utils import LOGGER, yaml_load
+from ultralytics.utils import DEFAULT_CFG_DICT, DEFAULT_SOL_DICT, LOGGER
 from ultralytics.utils.checks import check_imshow, check_requirements
 
-DEFAULT_SOL_CFG_PATH = Path(__file__).resolve().parents[1] / "cfg/solutions/default.yaml"
-
 
 class BaseSolution:
     """
@@ -55,15 +52,18 @@ class BaseSolution:
         self.Point = Point
 
         # Load config and update with args
-        self.CFG = yaml_load(DEFAULT_SOL_CFG_PATH)
+        DEFAULT_SOL_DICT.update(kwargs)
-        self.CFG.update(kwargs)
+        DEFAULT_CFG_DICT.update(kwargs)
-        LOGGER.info(f"Ultralytics Solutions: ✅ {self.CFG}")
+        self.CFG = {**DEFAULT_SOL_DICT, **DEFAULT_CFG_DICT}
+        LOGGER.info(f"Ultralytics Solutions: ✅ {DEFAULT_SOL_DICT}")
 
         self.region = self.CFG["region"]  # Store region data for other classes usage
-        self.line_width = self.CFG["line_width"]  # Store line_width for usage
+        self.line_width = (
+            self.CFG["line_width"] if self.CFG["line_width"] is not None else 2
+        )  # Store line_width for usage
 
         # Load Model and store classes names
-        self.model = YOLO(self.CFG["model"])
+        self.model = YOLO(self.CFG["model"] if self.CFG["model"] else "yolov8n.pt")
         self.names = self.model.names
 
         # Initialize environment and region setup

ultralytics/utils/__init__.py

@@ -38,6 +38,7 @@ FILE = Path(__file__).resolve()
 ROOT = FILE.parents[1]  # YOLO
 ASSETS = ROOT / "assets"  # default images
 DEFAULT_CFG_PATH = ROOT / "cfg/default.yaml"
+DEFAULT_SOL_CFG_PATH = ROOT / "cfg/solutions/default.yaml"  # Ultralytics solutions yaml path
 NUM_THREADS = min(8, max(1, os.cpu_count() - 1))  # number of YOLO multiprocessing threads
 AUTOINSTALL = str(os.getenv("YOLO_AUTOINSTALL", True)).lower() == "true"  # global auto-install mode
 VERBOSE = str(os.getenv("YOLO_VERBOSE", True)).lower() == "true"  # global verbose mode
@@ -508,6 +509,7 @@ def yaml_print(yaml_file: Union[str, Path, dict]) -> None:
 
 # Default configuration
 DEFAULT_CFG_DICT = yaml_load(DEFAULT_CFG_PATH)
+DEFAULT_SOL_DICT = yaml_load(DEFAULT_SOL_CFG_PATH)  # Ultralytics solutions configuration
 for k, v in DEFAULT_CFG_DICT.items():
     if isinstance(v, str) and v.lower() == "none":
         DEFAULT_CFG_DICT[k] = None
@@ -566,12 +568,16 @@ def is_kaggle():
 
 def is_jupyter():
     """
-    Check if the current script is running inside a Jupyter Notebook. Verified on Colab, Jupyterlab, Kaggle, Paperspace.
+    Check if the current script is running inside a Jupyter Notebook.
 
     Returns:
         (bool): True if running inside a Jupyter Notebook, False otherwise.
+
+    Note:
+        - Only works on Colab and Kaggle, other environments like Jupyterlab and Paperspace are not reliably detectable.
+        - "get_ipython" in globals() method suffers false positives when IPython package installed manually.
     """
-    return "get_ipython" in globals()
+    return IS_COLAB or IS_KAGGLE
 
 
 def is_docker() -> bool:
@@ -799,10 +805,10 @@ def get_user_config_dir(sub_dir="Ultralytics"):
 PROC_DEVICE_MODEL = read_device_model()  # is_jetson() and is_raspberrypi() depend on this constant
 ONLINE = is_online()
 IS_COLAB = is_colab()
+IS_KAGGLE = is_kaggle()
 IS_DOCKER = is_docker()
 IS_JETSON = is_jetson()
 IS_JUPYTER = is_jupyter()
-IS_KAGGLE = is_kaggle()
 IS_PIP_PACKAGE = is_pip_package()
 IS_RASPBERRYPI = is_raspberrypi()
 GIT_DIR = get_git_dir()
@@ -1193,7 +1199,7 @@ class SettingsManager(JSONDict):
             "neptune": True,  # Neptune integration
             "raytune": True,  # Ray Tune integration
             "tensorboard": True,  # TensorBoard logging
-            "wandb": True,  # Weights & Biases logging
+            "wandb": False,  # Weights & Biases logging
             "vscode_msg": True,  # VSCode messaging
         }
 

ultralytics/utils/checks.py

@@ -335,7 +335,7 @@ def check_font(font="Arial.ttf"):
         return file
 
 
-def check_python(minimum: str = "3.8.0", hard: bool = True, verbose: bool = True) -> bool:
+def check_python(minimum: str = "3.8.0", hard: bool = True, verbose: bool = False) -> bool:
     """
     Check current python version against the required minimum version.