Recently I read about this article and was amazed by the app’s ability to detect 3D objects in real time, so I decided to give it a try.
I followed the docker installation to build the mediapipe first, then build the objectdetection3d for android. This is the shoe detection only. After the docker build finish I need to copy the apk from the docker to my local machine via
docker cp mediapipe:mediapipe/bazel-bin/mediapipe/examples/android/src/java/com/google/mediapipe/apps/objectdetection3d/objectdetection3d.apk /dest_dir
I only have a MBP so I spent some time figuring out how to transfer the apk to android phone. Eventually I downloaded the google drive on the phone and install the apk from there. Here are the results, the poses are very accurate, but the speed is very slow. I used redmi note 7 and the video mode turned into picture mode…
The chair version is provided at objectron. This version seems slower than the shoe version, and the results are not as good
The paper is also released, and it says the EPnP details are in the supplementary. I am curious about how the authors define the 3D points of each object in the EPnP.