Adel Ahmadyan

1st stage = 2D object detector with SSD 2nd stage = keypoint prediction + EPNP The model above takes a 224x224 crop (containing the object) and produces 9 2D keypoints...

- We have not released the training code, but we did release the pre-trained models (both in Mediapipe, as well as the Objectron bucket on GCS). - The python evaluation...

The near and far plane are optimized per frame, and they are available via the projection_matrix, specifically, proj[2, 2] and proj[2, 3]  ) Which after solving should...

Thanks for reporting this. We ran a privacy filter on the bike data to blur all the faces & license plates. Sometime the filter does not work as expected and...

Can you be more specific as to which camera tracking you are referring to? We have open-sourced our own camera tracking library in Mediapipe [here](https://google.github.io/mediapipe/solutions/box_tracking.html#object-detection-and-tracking).

To start, you'll need annotated data. There are annotated cereal boxes in the dataset, but you'll need to collect your own data and annotate it, then you can train the...

We haven't released the training code for the models yet, so you have to implement your own model.

A good starting point would be [Tensorflow tutorials](https://www.tensorflow.org/tutorials), next you can look at the source code of relevant models on Github. https://paperswithcode.com/task/6d-pose-estimation

images are recorded in portrait mode, but the camera poses are in landscape.

Some of the models (shoe, chair, camera, and cup) can be downloaded from [Mediapipe website](https://google.github.io/mediapipe/solutions/objectron). Full set of models (along python solutions) are to be released later this month.