add point model

docs/user_guide/13_annotation.md

@@ -73,7 +73,7 @@ Yes! Object detection models use non-annotated areas of an image as negative dat
 
 1. **Select Important Images**: Duplicate backgrounds or objects contribute little to model generalization. Focus on gathering a wide variety of object appearances.
 2. **Avoid Over-splitting Labels**: Often, using a superclass for detection followed by a separate model for classification is more effective. See the [`CropModel`](03_cropmodels.md) for an example.
-3. **Balance Accuracy and Practicality**: Depending on the goal (e.g., object counting or detection), keypoints can sometimes be used instead of precise boxes to simplify the process.
+3. **Balance Accuracy and Practicality**: Depending on the goal (e.g., object counting or detection), point/keypoint annotations can sometimes be used instead of precise boxes to simplify the process.
 
 ## Quick Video on Annotating Images
 

pyproject.toml

@@ -51,6 +51,7 @@ dependencies = [
     "pyyaml>=5.1.0",
     "rasterio",
     "safetensors<0.6.0",
+    "scikit-image>=0.25.2",
     "setuptools",
     "shapely>2.0.0",
     "slidingwindow",

src/deepforest/conf/config.yaml

@@ -142,3 +142,8 @@ cropmodel:
     normalize:
     # Number of pixels to expand bbox crop windows for better prediction context.
     expand: 0
+
+point:
+  score_integration_radius: 5
+  nms_distance_thresh: 5.0
+  distance_threshold: 10

src/deepforest/conf/point.yaml

@@ -0,0 +1,22 @@
+architecture: treeformer
+
+model:
+  name: 'weecology/deepforest-tree-point'
+  revision: 'main'
+
+# This model was trained on 10 cm/px data and has been tested
+# at resolutions around 5-20 cm/px. If your data is significantly
+# different, for example < 5 cm/px drone data, we suggest first
+# downsampling your orthomosaics.
+patch_size: 512
+patch_overlap: 0.1
+
+# Setting a threshold below 0.3 is likely to include a lot of
+# noisy/low-confidence predictions.
+score_thresh: 0.3
+
+point:
+  backbone: 'pvt_v2_b3'
+  score_integration_radius: 5
+  nms_distance_thresh: 5.0
+  distance_threshold: 10.0

src/deepforest/conf/schema.py

@@ -41,7 +41,7 @@ class SchedulerConfig:
     """Set the type of scheduler, by default DeepForest uses a stepped learning
     function reducing at "milestones" during training."""
 
-    type: str | None = "StepLR"
+    type: str | None = "stepLR"
     params: SchedulerParamsConfig = field(default_factory=SchedulerParamsConfig)
 
 
@@ -57,8 +57,10 @@ class OptimizerConfig:
 
 @dataclass
 class TrainConfig:
-    """Main training configuration. The CSV file and root directory are
-    required to specify the location of the training dataset.
+    """Main training configuration.
+
+    The CSV file and root directory are required to specify the location
+    of the training dataset.
 
     The default learning rate may need to be changed for certain
     architectures, such as transformers-based models which sometimes
@@ -84,8 +86,10 @@ class TrainConfig:
 
 @dataclass
 class ValidationConfig:
-    """Main validation configuration. As with training data, it's required that
-    you set a CSV file and root directory.
+    """Main validation configuration.
+
+    As with training data, it's required that you set a CSV file and
+    root directory.
 
     Validation during training is important to identify if the model has
     converged or is overfitting.
@@ -125,11 +129,22 @@ class CropModelConfig:
     expand: int = 0
 
 
+@dataclass
+class PointConfig:
+    """Configuration for point models."""
+
+    backbone: str = "pvt_v2_b3"
+    score_integration_radius: int = 5
+    nms_distance_thresh: float = 5.0
+    distance_threshold: float = 10.0
+
+
 @dataclass
 class Config:
-    """General DeepForest configuration. Some parameters here are shared
-    between dataloaders, for example the batch size, accelerator and number of
-    workers.
+    """General DeepForest configuration.
+
+    Some parameters here are shared between dataloaders, for example the
+    batch size, accelerator and number of workers.
 
     Here we also set the architecture, which can be one of "retinanet"
     or "DeformableDetr" currently. If you modify the number of classes
@@ -173,3 +188,4 @@ class Config:
     validation: ValidationConfig = field(default_factory=ValidationConfig)
     predict: PredictConfig = field(default_factory=PredictConfig)
     cropmodel: CropModelConfig = field(default_factory=CropModelConfig)
+    point: PointConfig = field(default_factory=PointConfig)

src/deepforest/datasets/training.py

@@ -90,9 +90,10 @@ def _validate_labels(self) -> None:
 
     @abstractmethod
     def _validate_coordinates(self) -> None:
-        """Validate geometries in the annotation data. Must be overidden by
-        child classes to implement task-specific checks (e.g., boxes vs
-        points).
+        """Validate geometries in the annotation data.
+
+        Must be overidden by child classes to implement task-specific
+        checks (e.g., boxes vs points).
 
         Should raise ValueError with details if any invalid geometries
         are found.
@@ -285,8 +286,8 @@ def __getitem__(self, idx) -> tuple:
         return image, targets, self.image_names[idx]
 
 
-class KeypointDataset(TrainingDataset):
-    """Dataset for keypoint detection tasks."""
+class PointDataset(TrainingDataset):
+    """Dataset for point detection tasks."""
 
     _data_keys = [DataKey.IMAGE, DataKey.KEYPOINTS]