Configuration reference

All settings live in data/interim/pipeline.yaml. Values marked TUNABLE are the first candidates for adjustment when adapting the pipeline to a new camera setup.

image

Key	Type	Default	Description
margin_px	int	272	TUNABLE. Height (in pixels) of the camera metadata strip at the bottom of each image. Cropped before detection. Measure the strip height for your specific camera model.
input_extension	str	".jpg"	File extension glob used by process_directory. Case-insensitive (both .jpg and .JPG are always matched).

detector

Key	Type	Default	Description
model	str	"yolo11n.pt"	TUNABLE. YOLO checkpoint filename. Larger variants (yolo11s.pt, yolo11m.pt) are more accurate at the cost of speed.
model_dir	str	"models/YOLO"	Directory containing the YOLO checkpoint, relative to project root.
bird_class_id	int	14	COCO class index for "bird". Should not need changing unless using a custom YOLO model.
confidence_threshold	float	0.3	TUNABLE. Minimum YOLO detection score. Lower values catch more birds (higher recall, lower precision).
slice_size	int or null	1280	TUNABLE. SAHI tile size in pixels. Set to null to disable tiling (faster but may miss small subjects).
overlap_ratio	float	0.2	TUNABLE. Fractional overlap between adjacent SAHI tiles. Larger values reduce missed detections at tile boundaries.
nms_iou_threshold	float	0.5	TUNABLE. IoU threshold for SAHI post-NMS tile merge. Lower values keep more boxes — reduce to ~0.3 if two nearby birds are incorrectly merged.

classifier

Key	Type	Default	Description
model	str	"hummingbird_classifier.pt"	Classifier checkpoint filename.
model_dir	str	"models"	Directory containing the classifier checkpoint.
image_size	int	224	Input size fed to EfficientNetV2-S. Do not change unless you retrain with a different input resolution.
hummingbird_label	int	1	Numeric label assigned to positive (hummingbird) predictions.
confidence_threshold	float	0.5	TUNABLE. Minimum sigmoid probability to classify as hummingbird.

training

Used by scripts/train_classifier.py and overridable via CLI flags.

Key	Type	Default	Description
base_model	str	"efficientnet_v2_s"	TUNABLE. Backbone architecture.
epochs	int	30	TUNABLE. Total training epochs.
batch_size	int	32	TUNABLE. Batch size. Scale LR linearly when changing (new_lr = old_lr * new_bs / old_bs).
learning_rate	float	0.0003	TUNABLE. Head-phase learning rate.
weight_decay	float	0.0001	L2 regularisation.
val_split	float	0.2	Fraction of data reserved for validation (when val_dir is not supplied).
early_stopping_patience	int	5	Stop training if val loss does not improve for this many epochs.

deduplication

Key	Type	Default	Description
window_seconds	int	18	TUNABLE. Records within this time window are considered duplicates of the same event.

preprocessing

Key	Type	Default	Description
pad_fraction	float	0.10	TUNABLE. Fractional padding added around each detected bounding box before cropping the bird patch. 0.10 adds 10 % padding on each side.