# AR watches try-on ## Proposal 3 months ## Breakdown - Research — 60 hours - Hand 3D landmarks estimation — 80 hours - Arm segmentation — 80 hours - Pose tracking — 40 hours - Overlay 3D object — 40 hours - Port computer vision models to iOS — 40 hours - iOS SDK — 40 hours - Demo app — 40 hours - Documentation — 16 hours - Project management — 24 hours ## Approach There are two approaches. First one is to print a marker, that can be recognized by the app to overlay 3D-model, and place it on the wrist. This is how https://ar-watches.com did it. This is not convenient for users, and their app has 1 star rating at the app store. We do not recommend to go this way. ![](https://i.imgur.com/1wry1yi.png) The second one is to do it like AR apps with the face masks. They recognize face landmarks and then overlay mask image. ![](https://i.imgur.com/svXBHP0.jpg) ![](https://i.imgur.com/SUEexH8.png) ### Hand landmarks We need to recognize 3D pose of the wrist, but it will be not precise in the close-up photo, as unlike face, it does not have any keypoints, like nose, eyes, mouth. But if we will recognize landmarks of the hand first, then we can estimate 3D-pose of the wrist much better. ![](https://i.imgur.com/7HM0NYE.jpg) ### Arm segmentation The hand will give us 3D-direction of the wrist, and not a 3D-model. We will apply hand segmentation model to know where wrist starts and ends. Then we can place watches correctly. ![](https://i.imgur.com/YRF1DvV.png) ### Pose tracking These algorithms will have some error and 3D model will jitter if we won't apply filtering. Pose tracking will help to smooth animation and minimize 3D model shaking. ### 3D model overlay We will use open-source 3D model of the watch. It will be rendered on the photo from a camera. 3D position and cutting planes will be calculated based on the algorithms above. Other 3D models of watches can be added later and their modeling is out of current scope. ![](https://i.imgur.com/cNVEDMu.png) ### Demo app Demo app will not have any UI, just will show video stream from the camera with overlayed watches on wrist. It will work on iOS. ### Documentation There will be 2 documents: to describe the computer vision pipeline and to describe requirements for 3D models of watches. --- ###### tags: `proposal`