Hi everyone.

The following paragraph is extracted from your paper: "We used this tool to estimate the amount of human corrections needed to be “perfect” (most needed no corrections, and minor corrections were under 10 pixels, checked on n=600 frames) and the largest adjustments were typically in the extremities, such as the tail and ankles. The proportion of large outliers where the adjustment was over 100 mm in magnitude was only 0.71%".

Are these 600 frames and their corrected annotations are available anywhere? Alternatively, are any hand-checked 3D GT frames/annotations available anywhere? Thanks in advance!

Hi everyone.

Thank you for your work, and for making your data and results freely available.

I'm trying to understand the difference between the 2D annotations stored in the dlc_pw and fte_pw folders in your Dropbox repository: are the 2D annotations in the .csv files in fte_pw obtained by re-projecting the 3D poses to 2D, and are the annotations in the .pickle files in dlc_pw generated by the DeepLabCut model which are used as GT (to compare with the re-projected 2D annotations)?

Thanks in advance :)

When args.end_frame is set -1, the start_frame and end_frame are automatically defined with the following process:

max_idx = filtered_points_2d_df['frame'].max() + 1
for i in range(max_idx):
    if num_marker(i) == len(target_markers):
        start_frame = i
        break
for i in range(max_idx, 0, -1):
    if num_marker(i) == len(target_markers):
        end_frame = i
        break

line 812 on all_optimizations.py.

• run.sh ... make and run the docker environment
• count_frames.py ... display the number of frames
• FTE.py ... run FTE with CLI

You can run these like this

$ python FTE.py --data data/2019_03_03/menya/flick
$ python frame_count.py --video /data/2019_03_03/menya/flick/fte/fte.avi

readme updates

nb_conda file missing from yaml; also @DJoska @rickyjericevich is deeplabcut really required? also why a "_windows" version of the yaml?

Hi everyone.

I was hoping someone might be able to help me with this: does the DeepLabCut feature extractor used (with ResNet152 backbone) estimate the cheetah's pose from an entire frame, or is the body first localised by calculating its bounding box, before performing feature extraction?

Thanks in advance :)

Hello, authors: Thank you for your contributions. Here are some questions that need your help.

• First, is the 3D pose provided by AcinoSet in the world coordinate system (i.e., "fte.pickle"->traj_data["positions"])? Is the measurement in meters?
• Second, can the 3D pose in the world coordinate system be converted to the camera coordinate system through the "R, T" parameter in "n_cam_scene_sba.json"?

Thanks.

Initially when I created all_optimizations.py, I made sure that the functions for each optimization therein held the exact same code as it's corresponding notebook. Eg: the fte() func had the exact same code as FTE.ipynb. This made it relatively easy to transfer any changes made in the notebooks to all_optimizations.py and vice versa.

The person(s) who recently made changes to all_optimizations.py did not transfer their changes to the notebooks, so the code in all_optimizations.py has become substantially different to the code in the notebooks (and rather messy imo). This makes it difficult and time consuming to add any new changes made in the notebooks to all_optimizations.py and vice versa.

As of right now, all_optimizations.py will probably produce different reconstruction results compared to the notebooks, and that's a problem. This must be fixed before we can merge the improvements from develop into main

I think the best (long-term) solution is to break up the code in each notebook into smaller, more manageable functions and place them in their own module. For example, there'll be one module for all the FTE code and one for EKF and so on. Perhaps the FTE module will have functions like plot_redescending_cost(), initialize_pyomo_model(), define_pyomo_constraints() or something similar.

Thoughts?