Yaw_rate not changing from 0

[qzramiz]

On passing any type of image as an input there is no change in the yaw_rate from 0. It always selects the 87th bin and whose value lies the the range of [-small_angle +small_angle] so the resulting value becomes zero. I am passing three images consecutively as an input which are appended in latest_frames in the function observe_a_frame in wrapper.py. Is there any further changes required to get these values correctly. (This issue is only with the yaw_rate velocity is changing)

Thanks

[qzramiz]

waiting for your response.

[gy20073]

I am working on fixing this problem.

On reason might be that there is some domain shift between the BDD-V training dataset and your testing sequence. A simple fix would be redefining the bins. I will try to finish that and upload the model.

On Oct 22, 2017, at 2:59 AM, qzramiz notifications@github.com wrote:

waiting for your response.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/gy20073/BDD_Driving_Model/issues/5#issuecomment-338465567, or mute the thread https://github.com/notifications/unsubscribe-auth/AFv2jLvWXR8ispXlRwXbXXqVpA0zdG8sks5suxINgaJpZM4P2gRD.

[qzramiz]

It turns out that network is always giving the maximum logit value for 87th bin. what so ever is the input.

Outputs are mentioned below for the frames I've sent.

two best values for yaw rate: [88 87] [ 3.79977036 5.77287865] values in continuous map custom yaw & velocity [ 5.77287865] [ 3.02293372] [array([87]), array([39])] Current image : 84 Returned_values [[ 0. 2.82691396]] yaw, speed

two best values for yaw rate : [88 87] [ 4.09211683 6.03149223] values in continuous map custom yaw & velocity [ 6.03149223] [ 2.69692183] [array([87]), array([49])] Current image : 87 Returned_values [[ 0. 3.86596027]] yaw, speed

two best values for yaw rate : [88 87] [ 2.7101841 4.79574108] values in continuous map custom yaw & velocity[ 4.79574108] [ 3.49063468] [array([87]), array([33])] Current image : 90 Returned_values [[ 0. 2.19807248]] yaw, speed

two best values for yaw rate : [88 87] [ 2.81640673 4.89151764] values in continuous map custom yaw & velocity [ 4.89151764] [ 3.00964713] [array([87]), array([49])] Current image : 93 Returned_values [[ 0. 3.86596027]] yaw, speed

two best values for yaw rate: [88 87] [ 3.79965591 5.76331711] values in continuous map custom yaw & velocity [ 5.76331711] [ 2.62672067] [array([87]), array([61])] Current image : 96 Returned_values [[ 0. 5.09703919]] yaw, speed

two best values for yaw rate: [88 87] [ 3.94857597 5.97929001] values in continuous map custom yaw & velocity [ 5.97929001] [ 2.73023868] [array([87]), array([49])] Current image : 99 Returned_values [[ 0. 3.86596027]] yaw, speed

[qzramiz]

These frames are taken from KITTI dataset. Here are some frames that are feed as input. (It shows similar behavior on other datasets as well)

I hope this will help you in solving the issue. :)

[Shaswat27]

@qzramiz I am facing the same issue on KITTI sequences. Did you manage to solve the issue by any chance?

[merryHunter]

@gy20073 May I ask if there is any fix for this issue?

[merryHunter]

Hi @qzramiz, as I understand it's a bug in the code or model, what do you think?

[xuhuazhe]

There might be huge domain shift from BDD 100K to KITTI. Is it still the same issue when you use BDD 100K data?