search

AlextheEngineer / Ego3DHands

This repository is for the newest Ego3DHand dataset.
57 stars 7 forks source link

have you tested your model on real data? #2

Closed liwenssss closed 4 years ago

liwenssss commented 4 years ago

Hi, I use your data to train a simple model.But I found it worked worse on the real image data.Have you tested your model on the real image data?Can you give any advice?

AlextheEngineer commented 4 years ago

Thanks for asking! Generally, there is a domain gap between synthetic and real data. The differences between low-level features make it hard for CNNs to recognize. If you need to test on real images, you could try experimenting with FreiHand, CMU Panoptic or other datasets with real hand images. Note that annotations for real hand images are much more challenging, therefore such data is more limited in variety and quantity. Also, domain gap might still exist due to variables like pose space, lighting, view angle, etc. We are currently annotating a new dataset with egocentric two-hand real images and it will be ready later this year. Let me know if you have more questions!

From: liwenssss notifications@github.com Sent: Tuesday, June 16, 2020 1:15 AM To: AlextheEngineer/Ego3DHands Ego3DHands@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: [AlextheEngineer/Ego3DHands] have you tested your model on real data? (#2)

Hi, I use your data to train a simple model.But I found it worked worse on the real image data.Have you tested your model on the real image data?Can you give any advice?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/AlextheEngineer/Ego3DHands/issues/2, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ACV2RTHWLLWJIGEXXVNAE6TRW4LZ5ANCNFSM4N7KRIVA.

liwenssss commented 4 years ago

Nice work and I really look forward to it!

发自我的iPhone

------------------ Original ------------------ From: AlextheEngineer <notifications@github.com> Date: Wed,Jun 17,2020 3:22 AM To: AlextheEngineer/Ego3DHands <Ego3DHands@noreply.github.com> Cc: liwenssss <573014453@qq.com>, Author <author@noreply.github.com> Subject: Re: [AlextheEngineer/Ego3DHands] have you tested your model on real data? (#2)

Thanks for asking! Generally, there is a domain gap between synthetic and real data. The differences between low-level features make it hard for CNNs to recognize. If you need to test on real images, you could try experimenting with FreiHand or other datasets with real hand images. Although domain gap might still exist due to variables like pose space, image quality, lighting, etc. We are currently annotating a new dataset with egocentric two-hand real images and it will be ready later this year. Let me know if you have more questions!

From: liwenssss <notifications@github.com> Sent: Tuesday, June 16, 2020 1:15 AM To: AlextheEngineer/Ego3DHands <Ego3DHands@noreply.github.com> Cc: Subscribed <subscribed@noreply.github.com> Subject: [AlextheEngineer/Ego3DHands] have you tested your model on real data? (#2)

Hi, I use your data to train a simple model.But I found it worked worse on the real image data.Have you tested your model on the real image data?Can you give any advice?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<https://github.com/AlextheEngineer/Ego3DHands/issues/2&gt;, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ACV2RTHWLLWJIGEXXVNAE6TRW4LZ5ANCNFSM4N7KRIVA&gt;.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.