sahi 0.8.3

A vision library for performing sliced inference on large images/small objects

SAHI: Slicing Aided Hyper Inference

A lightweight vision library for performing large scale object detection & instance segmentation

Overview

Object detection and instance segmentation are by far the most important fields of applications in Computer Vision. However, detection of small objects and inference on large images are still major issues in practical usage. Here comes the SAHI to help developers overcome these real-world problems with many vision utilities.

Command	Description
predict	perform sliced/standard prediction using any yolov5/mmdet model
predict-fiftyone	perform sliced/standard prediction using any yolov5/mmdet model and explore results in fiftyone app
coco slice	automatically slice COCO annotation and image files
coco fiftyone	explore multiple prediction results on your COCO dataset with fiftyone ui ordered by number of misdetections
coco evaluate	evaluate classwise COCO AP and AR for given predictions and ground truth
coco analyse	calcualate and export many detection and segmentation error margin plots
coco yolov5	automatically convert any COCO dataset to yolov5 format

Getting Started

Blogpost

Check the official SAHI blog post.

Installation

• Install sahi using pip:

pip install sahi

• On Windows, Shapely needs to be installed via Conda:

conda install -c conda-forge shapely

• Install your desired version of pytorch and torchvision:

pip install torch torchvision

• Install your desired detection framework (such as mmdet or yolov5):

pip install mmdet mmcv

pip install yolov5

Usage

From Python:

• Sliced inference:

result = get_sliced_prediction(
    image,
    detection_model,
    slice_height = 256,
    slice_width = 256,
    overlap_height_ratio = 0.2,
    overlap_width_ratio = 0.2
)

Check YOLOv5 + SAHI demo:

Check MMDetection + SAHI demo:

• Slice an image:

from sahi.slicing import slice_image

slice_image_result = slice_image(
    image=image_path,
    output_file_name=output_file_name,
    output_dir=output_dir,
    slice_height=256,
    slice_width=256,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)

• Slice a coco formatted dataset:

from sahi.slicing import slice_coco

coco_dict, coco_path = slice_coco(
    coco_annotation_file_path=coco_annotation_file_path,
    image_dir=image_dir,
    slice_height=256,
    slice_width=256,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)

Refer to slicing notebook for detailed usage.

From CLI:

sahi predict --source image/file/or/folder --model_path path/to/model --model_config_path path/to/config

will perform sliced inference on default parameters and export the prediction visuals to runs/predict/exp folder.

You can specify sliced inference parameters as:

sahi predict --slice_width 256 --slice_height 256 --overlap_height_ratio 0.1 --overlap_width_ratio 0.1 --model_confidence_threshold 0.25 --source image/file/or/folder --model_path path/to/model --model_config_path path/to/config

• Specify postprocess type as --postprocess_type UNIONMERGE or --postprocess_type NMS to be applied over sliced predictions

• Specify postprocess match metric as --postprocess_match_metric IOS for intersection over smaller area or --match_metric IOU for intersection over union

• Specify postprocess match threshold as --postprocess_match_threshold 0.5

• Add --class_agnostic argument to ignore category ids of the predictions during postprocess (merging/nms)

• If you want to export prediction pickles and cropped predictions add --export_pickle and --export_crop arguments. If you want to change crop extension type, set it as --visual_export_format JPG.

• If you want to export prediction visuals, add --export_visual argument.

• By default, scripts apply both standard and sliced prediction (multi-stage inference). If you don't want to perform sliced prediction add --no_sliced_prediction argument. If you don't want to perform standard prediction add --no_standard_prediction argument.

• If you want to perform prediction using a COCO annotation file, provide COCO json path as add --dataset_json_path dataset.json and coco image folder as --source path/to/coco/image/folder, predictions will be exported as a coco json file to runs/predict/exp/results.json. Then you can use coco_evaluation command to calculate COCO evaluation results or coco_error_analysis command to calculate detailed COCO error plots.

Find detailed info on cli command usage (coco fiftyone, coco yolov5, coco evaluate, coco analyse) at CLI.md.

FiftyOne Utilities

Explore COCO dataset via FiftyOne app:

For supported version: pip install fiftyone>=0.11.1

from sahi.utils.fiftyone import launch_fiftyone_app

# launch fiftyone app:
session = launch_fiftyone_app(coco_image_dir, coco_json_path)

# close fiftyone app:
session.close()

Convert predictions to FiftyOne detection:

from sahi import get_sliced_prediction

# perform sliced prediction
result = get_sliced_prediction(
    image,
    detection_model,
    slice_height = 256,
    slice_width = 256,
    overlap_height_ratio = 0.2,
    overlap_width_ratio = 0.2
)

# convert detections into fiftyone detection format
fiftyone_detections = result.to_fiftyone_detections()

Explore detection results in Fiftyone UI:

sahi coco fifityone --image_dir dir/to/images --dataset_json_path dataset.json cocoresult1.json cocoresult2.json

will open a FiftyOne app that visualizes the given dataset and 2 detection results.

Specify IOU threshold for FP/TP by --iou_threshold 0.5 argument

COCO Utilities

COCO dataset creation:

• import required classes:

from sahi.utils.coco import Coco, CocoCategory, CocoImage, CocoAnnotation

• init Coco object:

coco = Coco()

• add categories starting from id 0:

coco.add_category(CocoCategory(id=0, name='human'))
coco.add_category(CocoCategory(id=1, name='vehicle'))

• create a coco image:

coco_image = CocoImage(file_name="image1.jpg", height=1080, width=1920)

• add annotations to coco image:

coco_image.add_annotation(
  CocoAnnotation(
    bbox=[x_min, y_min, width, height],
    category_id=0,
    category_name='human'
  )
)
coco_image.add_annotation(
  CocoAnnotation(
    bbox=[x_min, y_min, width, height],
    category_id=1,
    category_name='vehicle'
  )
)

• add coco image to Coco object:

coco.add_image(coco_image)

• after adding all images, convert coco object to coco json:

coco_json = coco.json

• you can export it as json file:

from sahi.utils.file import save_json

save_json(coco_json, "coco_dataset.json")

Convert COCO dataset to ultralytics/yolov5 format:

from sahi.utils.coco import Coco

# init Coco object
coco = Coco.from_coco_dict_or_path("coco.json", image_dir="coco_images/")

# export converted YoloV5 formatted dataset into given output_dir with a 85% train/15% val split
coco.export_as_yolov5(
  output_dir="output/folder/dir",
  train_split_rate=0.85
)

Get dataset stats:

from sahi.utils.coco import Coco

# init Coco object
coco = Coco.from_coco_dict_or_path("coco.json")

# get dataset stats
coco.stats
{
  'num_images': 6471,
  'num_annotations': 343204,
  'num_categories': 2,
  'num_negative_images': 0,
  'num_images_per_category': {'human': 5684, 'vehicle': 6323},
  'num_annotations_per_category': {'human': 106396, 'vehicle': 236808},
  'min_num_annotations_in_image': 1,
  'max_num_annotations_in_image': 902,
  'avg_num_annotations_in_image': 53.037243084530985,
  'min_annotation_area': 3,
  'max_annotation_area': 328640,
  'avg_annotation_area': 2448.405738278109,
  'min_annotation_area_per_category': {'human': 3, 'vehicle': 3},
  'max_annotation_area_per_category': {'human': 72670, 'vehicle': 328640},
}

Find detailed info on COCO utilities (yolov5 conversion, slicing, subsampling, filtering, merging, splitting) at COCO.md.

MOT Challenge Utilities

MOT Challenge formatted ground truth dataset creation:

• import required classes:

from sahi.utils.mot import MotAnnotation, MotFrame, MotVideo

• init video:

mot_video = MotVideo(name="sequence_name")

• init first frame:

mot_frame = MotFrame()

• add annotations to frame:

mot_frame.add_annotation(
  MotAnnotation(bbox=[x_min, y_min, width, height])
)

mot_frame.add_annotation(
  MotAnnotation(bbox=[x_min, y_min, width, height])
)

• add frame to video:

mot_video.add_frame(mot_frame)

• export in MOT challenge format:

mot_video.export(export_dir="mot_gt", type="gt")

• your MOT challenge formatted ground truth files are ready under mot_gt/sequence_name/ folder.

Find detailed info on MOT utilities (ground truth dataset creation, exporting tracker metrics in mot challenge format) at MOT.md.

Contributing

sahi library currently supports all YOLOv5 models and MMDetection models. Moreover, it is easy to add new frameworks.

All you need to do is, creating a new class in model.py that implements DetectionModel class. You can take the MMDetection wrapper or YOLOv5 wrapper as a reference.

Before opening a PR:

• Install required development packages:

pip install -U -e .[dev]

• Reformat with black and isort:

black . --config pyproject.toml
isort .

Contributors

Fatih Cagatay Akyon

Cemil Cengiz

Sinan Onur Altinuc

Kadir Sahin

Devrim Cavusoglu

Wei Ji