I've used ultralytics for YOLO once before, but just wondering if this was the industry/competition standard. Are there any other popular implementations? Is anyone creating a custom implementation?
", "messages": [ { "id": "3162242", "postDate": "03/29/2025 01:40:32", "content": "I've used ultralytics for YOLO once before, but just wondering if this was the industry/competition standard. Are there any other popular implementations? Is anyone creating a custom implementation?
", "rawMarkdown": "I've used ultralytics for YOLO once before, but just wondering if this was the industry/competition standard. Are there any other popular implementations? Is anyone creating a custom implementation?", "votes": null }, { "id": "3162418", "postDate": "03/29/2025 07:51:00", "content": "Here is timeline of evolution most popular branches/implementations of yolo:
\n\n\n.(2016).YOLOv1.Darknet
\n
\n …………|
\n .(2017).YOLOv2.Darknet
\n …………|
\n .(2018).YOLOv3.Darknet
\n …………|___________________________________________________
\n …………|………………………………………………………………………|
\n ….(2020).YOLOv4.Darknet(AlexeyAB)…..YOLOV4/YOLOR.Pytorch(WongKinYui).(2020)
\n ………………….________________________________________|___________
\n ………………….|………………………………………………………………………|
\n …YOLOV7.Pytorch(WongKinYui/AlexAB).(2023)….YOLOV5.Pytorch.(ultralytics).(2020)
\n …………………|……………………………………………………………..|___________________________________________________
\n …………………|……………………………………………………………..|…………………………….|…………………………………………..|
\n …YOLOV9.Pytorch(WongKinYui).(2024)………………..YOLO-NAS……..YOLOV8.(ultralytics).(2023?)…….YOLOV6.(2022)
\n …………………………………………………………………………………………………………………|___________________________________
\n …………………………………………………………………………………………………………………|………………………………………………….|
\n …………………………………………………………………………………YOLOV11.(ultralytics).(2024?)….YOLOV12.(sunsmarterjie).(2024)
Let me know if I missed something.
", "rawMarkdown": "Here is timeline of evolution most popular branches/implementations of yolo:\n\n>.(2016).YOLOv1.Darknet\n............|\n.(2017).YOLOv2.Darknet\n............|\n.(2018).YOLOv3.Darknet\n............|_______________________________________________________________\n............|.................................................................................|\n....(2020).YOLOv4.Darknet(AlexeyAB).....YOLOV4/YOLOR.Pytorch(WongKinYui).(2020)\n......................____________________________________________________|_______________________\n......................|.................................................................................|\n...YOLOV7.Pytorch(WongKinYui/AlexAB).(2023)....YOLOV5.Pytorch.(ultralytics).(2020)\n.....................|.......................................................................________|_____________________________________________________________\n.....................|.......................................................................|..................................|..................................................|\n...YOLOV9.Pytorch(WongKinYui).(2024)....................YOLO-NAS........YOLOV8.(ultralytics).(2023?).......YOLOV6.(2022)\n.................................................................................................................................|_______________________________________________\n.................................................................................................................................|..........................................................|\n.............................................................................................YOLOV11.(ultralytics).(2024?)....YOLOV12.(sunsmarterjie).(2024)\n\nLet me know if I missed something.", "votes": null }, { "id": "3167706", "postDate": "04/01/2025 18:59:25", "content": "You can think of different ways yourself. There are many approaches to solve the problem.
\nFor example template matching is a good idea.
\nActually I don't know why everyone are using YOLO. Maybe because the starter notebooks include YOLO implementation.
I'm trying out ultralytics's YOLO, but understanding its implementation is difficult and it's hard to customize.
\nI want to make modifications such as changing the number of input channels, but it's not easy.
\nI also want the implementation to follow the \"You Only Look Once\" concept.
try mmdetection - there you can customize whatever you need, to dive into mmdetection is pretty hard but once you do it you won't be limited only with ultralytics models
", "rawMarkdown": "try mmdetection - there you can customize whatever you need, to dive into mmdetection is pretty hard but once you do it you won't be limited only with ultralytics models", "votes": null }, { "id": "3170796", "postDate": "04/05/2025 02:51:05", "content": "You can try mmdetection to custom your model architecture, including yolo. But how to speed up its inference may be a problem. I just train a simple CenterNet with this library but it took about 9 hours to submit. :D
", "rawMarkdown": "You can try mmdetection to custom your model architecture, including yolo. But how to speed up its inference may be a problem. I just train a simple CenterNet with this library but it took about 9 hours to submit. :D", "votes": null }, { "id": "3171344", "postDate": "04/05/2025 15:34:41", "content": "I have an idea that I haven't tried yet. For inference, I plan to use only slices from 0 to 300, as the training data mostly contains flagellar motors within this range. However, I'm unsure how this will affect the LB score.
", "rawMarkdown": "I have an idea that I haven't tried yet. For inference, I plan to use only slices from 0 to 300, as the training data mostly contains flagellar motors within this range. However, I'm unsure how this will affect the LB score.", "votes": null } ], "comments": [ { "id": 3162418, "author_name": "fautei", "author_url": "", "post_date": "03/29/2025 07:51:00", "content": "Here is timeline of evolution most popular branches/implementations of yolo:
\n\n\n.(2016).YOLOv1.Darknet
\n
\n …………|
\n .(2017).YOLOv2.Darknet
\n …………|
\n .(2018).YOLOv3.Darknet
\n …………|___________________________________________________
\n …………|………………………………………………………………………|
\n ….(2020).YOLOv4.Darknet(AlexeyAB)…..YOLOV4/YOLOR.Pytorch(WongKinYui).(2020)
\n ………………….________________________________________|___________
\n ………………….|………………………………………………………………………|
\n …YOLOV7.Pytorch(WongKinYui/AlexAB).(2023)….YOLOV5.Pytorch.(ultralytics).(2020)
\n …………………|……………………………………………………………..|___________________________________________________
\n …………………|……………………………………………………………..|…………………………….|…………………………………………..|
\n …YOLOV9.Pytorch(WongKinYui).(2024)………………..YOLO-NAS……..YOLOV8.(ultralytics).(2023?)…….YOLOV6.(2022)
\n …………………………………………………………………………………………………………………|___________________________________
\n …………………………………………………………………………………………………………………|………………………………………………….|
\n …………………………………………………………………………………YOLOV11.(ultralytics).(2024?)….YOLOV12.(sunsmarterjie).(2024)
Let me know if I missed something.
", "votes": null, "replies": [] }, { "id": 3167706, "author_name": "vigneshwar472", "author_url": "", "post_date": "04/01/2025 18:59:25", "content": "You can think of different ways yourself. There are many approaches to solve the problem.
\nFor example template matching is a good idea.
\nActually I don't know why everyone are using YOLO. Maybe because the starter notebooks include YOLO implementation.
I'm trying out ultralytics's YOLO, but understanding its implementation is difficult and it's hard to customize.
\nI want to make modifications such as changing the number of input channels, but it's not easy.
\nI also want the implementation to follow the \"You Only Look Once\" concept.
try mmdetection - there you can customize whatever you need, to dive into mmdetection is pretty hard but once you do it you won't be limited only with ultralytics models
", "votes": null, "replies": [] } ] }, { "id": 3170796, "author_name": "i2nfinit3y", "author_url": "", "post_date": "04/05/2025 02:51:05", "content": "You can try mmdetection to custom your model architecture, including yolo. But how to speed up its inference may be a problem. I just train a simple CenterNet with this library but it took about 9 hours to submit. :D
", "votes": null, "replies": [ { "id": 3171344, "author_name": "rustambazarbayev", "author_url": "", "post_date": "04/05/2025 15:34:41", "content": "I have an idea that I haven't tried yet. For inference, I plan to use only slices from 0 to 300, as the training data mostly contains flagellar motors within this range. However, I'm unsure how this will affect the LB score.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3162242": "I've used ultralytics for YOLO once before, but just wondering if this was the industry/competition standard. Are there any other popular implementations? Is anyone creating a custom implementation?", "3162418": "Here is timeline of evolution most popular branches/implementations of yolo:\n\n>.(2016).YOLOv1.Darknet\n............|\n.(2017).YOLOv2.Darknet\n............|\n.(2018).YOLOv3.Darknet\n............|_______________________________________________________________\n............|.................................................................................|\n....(2020).YOLOv4.Darknet(AlexeyAB).....YOLOV4/YOLOR.Pytorch(WongKinYui).(2020)\n......................____________________________________________________|_______________________\n......................|.................................................................................|\n...YOLOV7.Pytorch(WongKinYui/AlexAB).(2023)....YOLOV5.Pytorch.(ultralytics).(2020)\n.....................|.......................................................................________|_____________________________________________________________\n.....................|.......................................................................|..................................|..................................................|\n...YOLOV9.Pytorch(WongKinYui).(2024)....................YOLO-NAS........YOLOV8.(ultralytics).(2023?).......YOLOV6.(2022)\n.................................................................................................................................|_______________________________________________\n.................................................................................................................................|..........................................................|\n.............................................................................................YOLOV11.(ultralytics).(2024?)....YOLOV12.(sunsmarterjie).(2024)\n\nLet me know if I missed something.", "3167706": "You can think of different ways yourself. There are many approaches to solve the problem.\nFor example template matching is a good idea.\nActually I don't know why everyone are using YOLO. Maybe because the starter notebooks include YOLO implementation.", "3169090": "I'm trying out ultralytics's YOLO, but understanding its implementation is difficult and it's hard to customize.\nI want to make modifications such as changing the number of input channels, but it's not easy.\nI also want the implementation to follow the \"You Only Look Once\" concept.", "3169421": "try mmdetection - there you can customize whatever you need, to dive into mmdetection is pretty hard but once you do it you won't be limited only with ultralytics models", "3170796": "You can try mmdetection to custom your model architecture, including yolo. But how to speed up its inference may be a problem. I just train a simple CenterNet with this library but it took about 9 hours to submit. :D", "3171344": "I have an idea that I haven't tried yet. For inference, I plan to use only slices from 0 to 300, as the training data mostly contains flagellar motors within this range. However, I'm unsure how this will affect the LB score." }, "source": "meta" }