{"id":2508,"date":"2021-10-20T09:00:30","date_gmt":"2021-10-20T01:00:30","guid":{"rendered":"https:\/\/blog.ailia.ai\/uncategorized\/yolox-object-detection-model-beyond-yolov5\/"},"modified":"2025-05-20T17:33:10","modified_gmt":"2025-05-20T09:33:10","slug":"yolox-object-detection-model-beyond-yolov5","status":"publish","type":"post","link":"https:\/\/blog.ailia.ai\/en\/tips-en\/yolox-object-detection-model-beyond-yolov5\/","title":{"rendered":"YOLOX : Object detection model exceeding YOLOv5"},"content":{"rendered":"\n

Overview<\/strong><\/h3>\n\n\n\n

YOLOX\u00a0<\/em>is a state-of-the-art object detection model released in August 2021, which combines performance beyond\u00a0YOLOv5<\/a>\u00a0with a permissive Apache license.<\/p>\n\n\n\n

\"\"
Source: https:\/\/github.com\/Megvii-BaseDetection\/YOLOX\/blob\/main\/assets\/logo.png<\/a><\/figcaption><\/figure>\n\n\n\n
\"\"
Source: https:\/\/github.com\/Megvii-BaseDetection\/YOLOX\/blob\/main\/assets\/demo.png<\/a><\/figcaption><\/figure>\n\n\n\n

<\/a>YOLOX: Exceeding YOLO Series in 2021<\/a><\/a><\/p>\n\n\n\n

GitHub – Megvii-BaseDetection\/YOLOX<\/a><\/p>\n\n\n\n

Architecture<\/strong><\/h3>\n\n\n\n

YOLOX <\/em>is an object detection model that is an anchor-free version of the conventional YOLO <\/em>and introduces decoupled head<\/em> and SimOTA<\/em>. This model was awarded first place of the Streaming Perception Challenge<\/em> at CVPR2021 Automatic Driving Workshop<\/em>.<\/p>\n\n\n\n

Since the existing YOLOv4<\/em><\/a> <\/em>and YOLOv5<\/em><\/a> <\/em>pipelines are over-optimized for the use of anchors, YOLOX has been improved with YOLOv3-SPP<\/em><\/a> as a baseline. YOLOv3-SPP<\/em> was updated to use the advanced YOLOv5<\/em> architecture that adopts an advanced CSPNet<\/em> backbone and an additional PAN head<\/em>.<\/p>\n\n\n\n

In object detection models, the tasks of classification <\/em>and regression <\/em>(calculation of bounding box positions) are performed simultaneously, which is known to cause conflicts and reduce accuracy. To solve this problem, the concept of decoupled head<\/em> was introduced. The conventional YOLO<\/em> series backbone and feature pyramids still use a classic coupled head,<\/em>but YOLOX <\/em>has been updated to use a decoupled head<\/em> and achieve higher accuracy.<\/p>\n\n\n\n

\"\"
Source: https:\/\/arxiv.org\/pdf\/2107.08430.pdf<\/a><\/figcaption><\/figure>\n\n\n\n

YOLOX<\/em> was trained on a dataset that was strongly augmented using Mosaic and Mixup strategies. The authors also use the advanced label assignment SimOTA<\/em>, a modified version of OTA<\/a>, to optimize loss.<\/p>\n\n\n\n

The contribution of each newly introduced tool is as follows.<\/p>\n\n\n\n

\"\"
Source: https:\/\/arxiv.org\/pdf\/2107.08430.pdf<\/a><\/figcaption><\/figure>\n\n\n\n

The benchmark results of\u00a0YOLOX\u00a0<\/em>are shown below.<\/p>\n\n\n\n

\"\"
Source: https:\/\/arxiv.org\/pdf\/2107.08430.pdf<\/a><\/figcaption><\/figure>\n\n\n\n

YOLOX model variants<\/strong><\/h3>\n\n\n\n

There are variations of\u00a0YOLOX<\/em>\u00a0split in two categories,\u00a0Standard Models<\/em>\u00a0for high precision and\u00a0Light Models<\/em>\u00a0for edge devices.<\/p>\n\n\n\n

\"\"
Source: https:\/\/github.com\/Megvii-BaseDetection\/YOLOX<\/a><\/figcaption><\/figure>\n\n\n\n

YOLOX performance<\/strong><\/h3>\n\n\n\n

Inference time and mAP50 was measured on validation set of COCO2017.\u00a0YOLOX-s<\/em>\u00a0is able to achieve the same accuracy as\u00a0YOLOv4\u00a0<\/em>with half processing time.<\/p>\n\n\n\n

\"\"
yolox\u306emAP50<\/figcaption><\/figure>\n\n\n\n
\"\"
yolox\u306e\u63a8\u8ad6\u6642\u9593<\/figcaption><\/figure>\n\n\n\n

The following repository and ailia SDK 1.2.8 were used to measure mAP and inference time.<\/p>\n\n\n\n

<\/a>GitHub – rafaelpadilla\/Object-Detection-Metrics<\/a><\/p>\n\n\n\n

CVPR2021 Automous Driving Workshop Streaming Perception Challenge<\/strong><\/h3>\n\n\n\n

The link below is the leaderboard of the\u00a0Streaming Perception Challenge<\/em>\u00a0atCVPR2021 Automatic Driving Workshop<\/em>, in which\u00a0YOLOX\u00a0<\/em>won the first place under the name\u00a0BaseDet<\/em>.<\/p>\n\n\n\n

<\/a>EvalAI: Evaluating state of the art in AI<\/a><\/p>\n\n\n\n

For this challenge,\u00a0<\/em>Argoverse 1.1<\/em><\/a>\u00a0dataset was used, which is the\u00a0Argoverse HD<\/em><\/a>dataset for automated driving with the addition of 2D bounding box annotations similar to the COCO dataset. The\u00a0Argoverse 1.1<\/em>\u00a0dataset contains 1,250,000 bounding boxes annotated using car frontal camera videos.<\/p>\n\n\n\n

\"\"
Source: https:\/\/www.cs.cmu.edu\/~mengtial\/proj\/streaming\/<\/a><\/figcaption><\/figure>\n\n\n\n

<\/a>Streaming Perception<\/a><\/p>\n\n\n\n

Usage<\/strong><\/h3>\n\n\n\n

YOLOX can be used with ailia SDK with the following command to detect object in the webcam video stream.<\/p>\n\n\n\n

$ python3 yolox.py -v 0<\/code><\/p>\n\n\n\n

By default,\u00a0YOLOX-s<\/em>\u00a0is used. Other models, including tiny models, can be used by using\u00a0-m<\/code>\u00a0option.<\/p>\n\n\n\n

<\/a>ailia-models\/object_detection\/yolk<\/a><\/p>\n\n\n\n

ax Inc.<\/a> has developed ailia SDK<\/a>, which enables cross-platform, GPU-based rapid inference.<\/p>\n\n\n\n

ax Inc. provides a wide range of services from consulting and model creation, to the development of AI-based applications and SDKs. Feel free to contact us<\/a> for any inquiry.<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Overview YOLOX\u00a0is a state-of-the-art object detection model released in August 2021, which combines performance beyond\u00a0YOLOv5\u00a0with a permissive Apache license. YOLOX: Exceeding YOLO Series in 2021 GitHub – Megvii-BaseDetection\/YOLOX Architecture YOLOX is an object detection model that is an anchor-free version of the conventional YOLO and introduces decoupled head and SimOTA. This model was awarded first place of the Streaming Perception Challenge at CVPR2021 Automatic Driving Workshop. Since the existing YOLOv4 and YOLOv5 pipelines are over-optimized for the use of anchors, YOLOX has been improved with YOLOv3-SPP as a baseline. YOLOv3-SPP was updated to use the advanced YOLOv5 architecture that adopts an advanced CSPNet backbone and an additional PAN head. In object detection models, the tasks of classification and regression (calculation of bounding box positions) are performed simultaneously, which is known to cause conflicts and reduce accuracy. […]<\/p>\n","protected":false},"author":5,"featured_media":2438,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[255],"tags":[266],"class_list":["post-2508","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tips-en","tag-ailiamodels-en"],"acf":[],"_links":{"self":[{"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/posts\/2508","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/comments?post=2508"}],"version-history":[{"count":1,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/posts\/2508\/revisions"}],"predecessor-version":[{"id":2510,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/posts\/2508\/revisions\/2510"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/media\/2438"}],"wp:attachment":[{"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/media?parent=2508"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/categories?post=2508"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.ailia.ai\/en\/wp-json\/wp\/v2\/tags?post=2508"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}