Configuration

CanopyRS pipelines are defined in YAML. A pipeline config lists the components to run and their parameters.

Pipeline config structure

components_configs:
  - tilerizer:
      tile_type: tile
      tile_size: 1777
      tile_overlap: 0.75
      ground_resolution: 0.045

  - detector: detectors/dino_swinL_multi_NQOS.yaml

  - aggregator:
      nms_algorithm: 'iou'
      score_threshold: 0.5
      nms_threshold: 0.7
      edge_band_buffer_percentage: 0.05
      scores_weights: { 'detector_score': 1.0 }

Each entry is either:

• Inline config — parameters specified directly in the pipeline YAML

• Reference — a path to a reusable component config, for example a detector in canopyrs/config/detectors/ or a segmenter in canopyrs/config/segmenters/. See our Model Zoo for the full list.

  Example reusable detector config (detectors/dino_swinL_multi_NQOS.yaml):

model: dino_detrex
architecture: dino/configs/dino-swin/dino_swin_large_384_5scale_36ep.py
checkpoint_path: 'https://huggingface.co/CanopyRS/dino-swin-l-384-multi-NQOS/resolve/main/model_best.pth'
batch_size: 1
box_predictions_per_image: 500
num_classes: 1

Key parameters

Tilerizer

Parameter	Description
tile_type	tile, tile_labeled, or polygon
tile_size	Tile size in pixels
tile_overlap	Overlap ratio between tiles (0–1)
ground_resolution	Target ground resolution in meters/pixel (cannot be set together with scale_factor)
scale_factor	Scale factor to resize tiles (cannot be set together with ground_resolution). A value of 1.0 keeps the source raster resolution as-is.

Detector

Parameter	Description
model	Model type identifier (e.g. dino_detrex, faster_rcnn_detectron2, deepforest)
architecture	Model architecture configuration path
checkpoint_path	Path or URL to pre-trained model weights
batch_size	Inference batch size
num_classes	Number of classes to detect (typically 1 for trees)
box_predictions_per_image	Maximum number of predictions per image
augmentation_image_size	Resize image size. A single int (e.g. 1024) for a fixed size, or a [min, max] list (e.g. [1024, 1777]) where images smaller than min are upscaled and images larger than max are downscaled

Segmenter

Parameter	Description
model	Model type identifier (e.g. sam2, sam3, detectree2)
architecture	Model architecture variant (e.g. l for large)
checkpoint_path	Path or URL to pre-trained model weights
image_batch_size	Number of images processed at once (sometimes ignored by models like SAM that rely exclusively on box_batch_size)
box_batch_size	Number of prompt boxes processed at once per image
augmentation_image_size	Resize image size. A single int (e.g. 1024) for a fixed size, or a [min, max] list (e.g. [1024, 1777]) where images smaller than min are upscaled and images larger than max are downscaled
pp_n_workers	Number of workers for parallel post-processing. Reduce this number if you have a limited amount of CPU cores.
pp_down_scale_masks_px	Downscale masks to this pixel size for more efficient post-processing
pp_simplify_tolerance	Polygon simplification tolerance
pp_remove_rings	Remove holes/rings from segmentation masks
pp_remove_small_geoms	Remove geometries smaller than this area threshold

Aggregator

Parameter	Description
nms_algorithm	iou (detection or segmentation) or ioa-disambiguate (segmentation only)
score_threshold	Minimum score to keep a detection
nms_threshold	IoU threshold for suppression
scores_weights	Weight per score column for combined scoring
scores_weighting_method	How to combine scores (e.g. weighted_geometric_mean)

Classifier

To be added.