sahi

What's Changed

• add url image read to read_image_as_pil by @fcakyon in https://github.com/obss/sahi/pull/283
• faster mmdet config download by @fcakyon in https://github.com/obss/sahi/pull/285
• add HuggingFace Spaces badge. by @devrimcavusoglu in https://github.com/obss/sahi/pull/282
• update to v0.8.11 by @fcakyon in https://github.com/obss/sahi/pull/284

Full Changelog: https://github.com/obss/sahi/compare/0.8.10...0.8.11

What's Changed

• improve mmdet config download util by @fcakyon in https://github.com/obss/sahi/pull/278
• update version to v0.8.10 by @fcakyon in https://github.com/obss/sahi/pull/279
• add zenodo doi and citation into readme by @fcakyon in https://github.com/obss/sahi/pull/280
• fix broken link in readme by @fcakyon in https://github.com/obss/sahi/pull/277
• fix mmcv installation in readme by @fcakyon in https://github.com/obss/sahi/pull/281

Full Changelog: https://github.com/obss/sahi/compare/0.8.9...0.8.10

What's Changed

• improve result visuals/plots by @fcakyon in https://github.com/obss/sahi/pull/271

• add get_coco_with_clipped_bboxes utility to coco class by @ssahinnkadir in https://github.com/obss/sahi/pull/264

• fix mmdet channel order by @fcakyon in https://github.com/obss/sahi/pull/266

• handle when mask prediction is empty by @fcakyon in https://github.com/obss/sahi/pull/269

• handle invalid annotation polygon by @fcakyon in https://github.com/obss/sahi/pull/268

• add missing shapely 2.0.0 support by @fcakyon in https://github.com/obss/sahi/pull/270

• update torch, torchvision, mmdet, mmcv versions in tests by @fcakyon in https://github.com/obss/sahi/pull/267

Full Changelog: https://github.com/obss/sahi/compare/0.8.8...0.8.9

enhancement

• add coco to yolov5 export utilty (#258)
• shapely 2.0.0 support, supported fiftyone versions (#257)
• update shapely to v1.8.0 (#259)

bugfixes

• fix a bug in coco upsampling/downsampling (#253)
• fix a bug in export_coco_as_yolov5 (#262)

ci

• update yolov5 version in tests (#256)

improvements

• update postprocess behaviour (#247)

testing

• update torch to 1.9.1 for workflows (#248)
• update mmdet version to 2.17.0 for testing (#250)
• add yolov5n weight download util (#251)
• improve test durations (#252)

bugfixes, enhancements

• handle when image name is very short in yolov5 conversion (#230)
• fix coco category creation without supercategory (#237)
• fix yolov5 conf threshold (#242)
• fix mmdet img size (#243)
• handle when mmdet config contains resize with keep_ratio=False (#244)
• update yolov5 version to 6.0.1 in ci tests (#245)

• handle predictions with negative bbox values (#225)
• fix remove_invalid_coco_results (#227)
• fix coco evaluation (#228)

Remove invalid coco results:

from sahi.utils.file import save_json
from sahi.utils.coco import remove_invalid_coco_results

# remove invalid predictions from COCO results JSON
coco_results = remove_invalid_coco_results("coco_result.json")

# export processed COCO results
save_json(coco_results, "fixed_coco_result.json")

# bonus: remove invalid predictions from COCO results JSON by giving COCO
# dataset path to also filter out bbox results exceeding image height&width
coco_results = remove_invalid_coco_results("coco_result.json", "coco_dataset.json")

• fix a bug in coco2fiftyone (#223)

• fix optional dependencies (#221)
• move conda builds to conda-forge (#219)

make repo compatible for conda-forge:

• Include requirements.txt in sdist (#217) thanks to @weiji14

new cli description:

new cli usage:

predict

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

predict-fiftyone

sahi predict-fiftyone --image_dir image/file/or/folder --dataset_json_path dataset.json --model_path path/to/model --model_config_path path/to/config

coco slice

sahi coco slice --image_dir dir/to/images --dataset_json_path dataset.json

coco fiftyone

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

coco evaluate

sahi coco evaluate --dataset_json_path dataset.json --result_json_path result.json

coco analyse

sahi coco analyse --dataset_json_path dataset.json --result_json_path result.json

coco yolov5

sahi coco yolov5 --image_dir dir/to/images --dataset_json_path dataset.json  --train_split 0.9

breaking changes in predict cli:

• config_path to model_config_path
• conf_thresh to model_confidence_threshold
• match_metric to postprocess_match_metric
• match_thresh to postprocess_match_threshold
• class_agnostic to postprocess_class_agnostic
• pickle to export_pickle, crop to export_crop
• novisual to export_visual
• no_sliced_pred to no_sliced_prediction
• no_standard_pred to no_standard_prediction
• coco_file to dataset_json_path

enhancements

• add script for inspecting cocoresults in fiftyone (#192)
• create symlinks for mot utils (#199)
• handle when iou_thrs is a single float (#196)
• Update coco_evaluation.py (#197)

bugfix

• fix coco_evaluation script (#195)
• fix a typo in object annotation (#194)

enhancements

• handle negative bbox coords, utilize image_size param (#188)

• add get_upsampled_coco utility to Coco (#189)

• add category and negative sample based coco up/sub-sampling (#191)

• Subsample COCO dataset file:

from sahi.utils.coco import Coco

# specify coco dataset path
coco_path = "coco.json"

# init Coco object
coco = Coco.from_coco_dict_or_path(coco_path)

# create a Coco object with 1/10 of total images
subsampled_coco = coco.get_subsampled_coco(subsample_ratio=10)

# export subsampled COCO dataset
save_json(subsampled_coco.json, "subsampled_coco.json")

# bonus: create a Coco object with 1/10 of total images that contain first category
subsampled_coco = coco.get_subsampled_coco(subsample_ratio=10, category_id=0)

# bonus2: create a Coco object with negative samples reduced to 1/10
subsampled_coco = coco.get_subsampled_coco(subsample_ratio=10, category_id=-1)

• Upsample COCO dataset file:

from sahi.utils.coco import Coco

# specify coco dataset path
coco_path = "coco.json"

# init Coco object
coco = Coco.from_coco_dict_or_path(coco_path)

# create a Coco object with each sample is repeated 10 times
upsampled_coco = coco.get_upsampled_coco(upsample_ratio=10)

# export upsampled COCO dataset
save_json(upsampled_coco.json, "upsampled_coco.json")

# bonus: create a Coco object with images that contain first category repeated 10 times
subsampled_coco = coco.get_subsampled_coco(upsample_ratio=10, category_id=0)

# bonus2: create a Coco object with negative samples upsampled by 10 times
upsampled_coco = coco.get_upsampled_coco(upsample_ratio=10, category_id=-1)

• add large tiff image reading support (#183)
• add concurrent slice export (#184)

• fix segmentation slicing (#174)

• refactor predict api (#170)

breaking changes

in predict and predict_fiftyone funtions:

• replaced model_name arg with model_type
• replaced model_parameters arg with model_path, model_config_path, model_confidence_threshold, model_device, model_category_mapping, model_category_remapping

in DetectionModel base class:

• replaced prediction_score_threshold arg with confidence_threshold

updated demo notebooks accordingly

• add input size parameter for inference (#169)

Example usages:

detection_model = Yolov5DetectionModel(
    model_path=yolov5_model_path,
    device="cpu", # or 'cuda'
)

result = detection_model.perform_inference(
    image,
    image_size=1280
)

result = get_prediction(
    "demo_data/small-vehicles1.jpeg",
    detection_model,
    image_size=1280
)

result = get_sliced_prediction(
    "demo_data/small-vehicles1.jpeg",
    detection_model,
    image_size=1280,
    slice_height = 256,
    slice_width = 256,
    overlap_height_ratio = 0.2,
    overlap_width_ratio = 0.2
)

• refactor slice_coco script (#165)
• make default for ignore_negative_samples, False (#166)

enhancement

• add coco_evaluation script, refactor coco_error_analysis script (#162)
  coco_evaluation.py script usage:

python scripts/coco_evaluation.py dataset.json results.json

will calculate coco evaluation and export them to given output folder directory.

If you want to specify mAP metric type, set it as --metric bbox mask.

If you want to also calculate classwise scores add --classwise argument.

If you want to specify max detections, set it as --proposal_nums 10 100 500.

If you want to specify a psecific IOU threshold, set it as --iou_thrs 0.5. Default includes 0.50:0.95 and 0.5 scores.

If you want to specify export directory, set it as --out_dir output/folder/directory .

coco_error_analysis.py script usage:

python scripts/coco_error_analysis.py dataset.json results.json

will calculate coco error plots and export them to given output folder directory.

If you want to specify mAP result type, set it as --types bbox mask.

If you want to export extra mAP bar plots and annotation area stats add --extraplots argument.

If you want to specify area regions, set it as --areas 1024 9216 10000000000.

If you want to specify export directory, set it as --out_dir output/folder/directory .

bugfixes

• prevent empty bbox coco json creation (#164)
• dont create mot info while when type='det' (#163)

breaking changes

• refactor predict (#161)
  By default, scripts apply both standard and sliced prediction (multi-stage inference). If you don't want to perform sliced prediction add --no_sliced_pred argument. If you don't want to perform standard prediction add --no_standard_pred argument.

• fix negative bbox coord error (#160)