Model catalog¶

ovkit serves a large set of permissively-licensed OpenVINO models from its mirror. Many are the same capability in several variants (different accuracy/speed, input resolution, INT8 quantization, or pruning), so this page is organized by function. Run ovkit list for the live list with descriptions, and ovkit info <name> for one model.

Capability aliases (recommended defaults)¶

Don’t want to pick among variants? Use a capability alias — a friendly name that points at a well-tested default. Model("face_detection") just works.

alias	→ default model	what it does
`detect`	`rtdetr_r50`	general object detection (COCO-80)
`face_detection`	`face_detection_0205`	detect faces
`person_detection`	`person_detection_0202`	detect people
`pedestrian_detection`	`person_detection_0202`	detect pedestrians (driving)
`vehicle_detection`	`vehicle_detection_0200`	detect vehicles
`text_detection`	`text_detection_0004`	detect text regions
`segment`	`pspnet_pytorch`	semantic segmentation (general scenes)
`road_segmentation`	`road_segmentation_adas_0001`	road / lane segmentation (driving)
`instance_segmentation`	`instance_segmentation_person_0007`	per-instance masks
`pose`	`human_pose_estimation_0007`	human body keypoints
`face_landmarks`	`landmarks_regression_retail_0009`	5-point face landmarks
`head_pose`	`head_pose_estimation_adas_0001`	head yaw/pitch/roll
`gaze`	`gaze_estimation_adas_0002`	gaze direction
`age_gender`	`age_gender_recognition_retail_0013`	age + gender
`emotion`	`emotions_recognition_retail_0003`	facial emotion
`face_reid`	`face_reidentification_retail_0095`	face embedding
`person_attributes`	`person_attributes_recognition_crossroad_0234`	clothing/attributes
`vehicle_attributes`	`vehicle_attributes_recognition_barrier_0042`	type + color
`classify`	`nfnet_f0`	image classification (ImageNet)
`image_retrieval`	`image_retrieval_0001`	image embedding
`super_resolution`	`single_image_super_resolution_1033`	3-4x upscaling
`text_recognition`	`text_recognition_0014`	read cropped text
`license_plate`	`vehicle_license_plate_detection_barrier_0106`	detect license plates
`qa`	`bert_small_uncased_whole_word_masking_squad_0002`	extractive Q&A
`translation`	`machine_translation_nar_en_de_0002`	EN→DE translation
`noise_suppression`	`noise_suppression_poconetlike_0001`	denoise speech
`time_series`	`time_series_forecasting_electricity_0001`	forecasting
`sound_classification`	`aclnet`	classify a sound clip (audio in — `model.infer()`)
`style_transfer`	`fast_neural_style_mosaic_onnx`	stylize an image (image out)
`background_matting`	`background_matting_mobilenetv2`	cut the subject out (frame + empty scene)
`llm`	`qwen25_1_5b_instruct`	chat LLM (genai)
`vlm`	`qwen3_vl_4b_instruct`	vision-language: image + question -> text
`stt`	`whisper_base`	speech-to-text (genai)

from ovkit import Model
Model("face_detection")("photo.jpg")     # uses face_detection_0205

Aliases live in src/ovkit/manifests/aliases.yaml — re-point any one in a line.

Registered models (one representative per capability)¶

The registry deliberately exposes one well-tested model per capability so pickers stay readable. Every variant (other accuracy/speed tiers, int8, sparse_NN, other resolutions) is still hosted on the mirror — see Variants below to surface one.

function	model
general object detection	`rtdetr_r50`
face detection	`face_detection_0205`
person detection	`person_detection_0202`
vehicle detection	`vehicle_detection_0200`
traffic combo (person+vehicle+bike)	`person_vehicle_bike_detection_2000`
text region detection	`text_detection_0004`
product detection	`product_detection_0001`
license plate detection	`vehicle_license_plate_detection_barrier_0106`
age + gender	`age_gender_recognition_retail_0013`
emotion	`emotions_recognition_retail_0003`
head pose	`head_pose_estimation_adas_0001`
face landmarks (5-pt)	`landmarks_regression_retail_0009`
face re-identification	`face_reidentification_retail_0095`
image classification	`resnet50_binary_0001`
person attributes	`person_attributes_recognition_crossroad_0234`
vehicle attributes	`vehicle_attributes_recognition_barrier_0042`
image embedding / retrieval	`image_retrieval_0001`
semantic segmentation	`road_segmentation_adas_0001`
instance segmentation	`instance_segmentation_person_0007`
human pose	`human_pose_estimation_0007`
gaze	`gaze_estimation_adas_0002`
scene text recognition	`text_recognition_0014`
handwritten Chinese OCR	`handwritten_simplified_chinese_recognition_0001`
sign language	`common_sign_language_0002`
weld defect detection	`weld_porosity_detection_0001`
question answering	`bert_small_uncased_whole_word_masking_squad_0002`
translation (en↔de, en↔ru)	`machine_translation_nar_{en_de, de_en, en_ru, ru_en}_0002`
speech noise suppression	`noise_suppression_poconetlike_0001`
time-series forecasting	`time_series_forecasting_electricity_0001`
super-resolution	`single_image_super_resolution_1033`
chat LLM	`tinyllama_chat`
speech-to-text	`whisper_base`

Variants¶

Suffix conventions on the mirror: numbered IDs = accuracy/speed tiers, int8 = quantized, sparse_NN = pruned, adas/retail = target domain. The mirror (leeyunjai/ovkit-models) hosts the full Apache-2.0 OMZ set; to surface a variant by name, add it to scripts/representatives.yaml and regenerate, or paste its entry into src/ovkit/manifests/omz.yaml:

python scripts/build_mirror.py --omz-intel --representatives \
    --emit-manifest src/ovkit/manifests/omz.yaml   # omit --representatives for ALL