PoseTrack18 results?

[ybai62868]

Hi, you write very good code and recently I am working on PoseTrack (including 3 tasks) too.
I use the msra's pytorch code (simple baseline) finetune the PoseTrack 2018 dataset and all parameters I use according to the paper.
Using gt_box on PoseTrack2018 dataset for Human Pose Estimation, I get the results which is worse than you reported:
my eval results:

Loading data
('# gt frames :', 3902)
('# pred frames:', 3902)
Evaluation of per-frame multi-person pose estimation
('saving results to', './out/total_AP_metrics.json')
Average Precision (AP) metric:
& Head & Shou & Elb & Wri & Hip & Knee & Ankl & Total\
& 49.4 & 89.2 & 83.9 & 76.1 & 81.0 & 80.3 & 74.6 & 74.6 \

Can you teach me how to finetune the model and I want to get the same results as you!
Thank you!
BTW, do you plan to implement the tracking code in this paper(simple baseline for human pose estimation and tracking)? I am doing this section ...

[mks0601]

To train a model on the PoseTrack 2018 dataset, I firstly trained a model on the COCO dataset. You can download the pre-trained model in the README. After that, renamed them from snapshot_140* to snapshot_0* and placed them at output/model_dump/PoseTrack/. Finally, I ran python train.py --gpu 0-1 --continue after changing dataset name to the PoseTrack in config.py. No hyperparameters changed from the config.py.

Unfortunately, I do not have a plan to implement tracking part :(

[ybai62868]

@mks0601 Thank you for your detailed replay！ After simply debug， I find that I inverted head top and head bottom horizontally by mistake. Which looks like very stupid. After a new training ...
I gets the results as follow:

Loading data
('# gt frames :', 3902)
('# pred frames:', 3902)
Evaluation of per-frame multi-person pose estimation
('saving results to', './out/total_AP_metrics.json')
Average Precision (AP) metric:
& Head & Shou & Elb & Wri & Hip & Knee & Ankl & Total\
& 88.0 & 89.2 & 84.1 & 76.2 & 81.1 & 80.5 & 75.0 & 82.4 \

[mks0601]

Sounds good!

[lxtGH]

@mks0601 Hi!! pose track dataset have 15 joints(2 joints are not annotated). How did you handle this?

[mks0601]

Just use pre-trained model on COCO, and set loss zero for not annotated l_ear and r_ear.

[Vhunon]

Where did you set the loss to zero in the code?

[mks0601]

I can't remember clearly, but
Please check L86 of main/gen_batch.py.
joints[:,2] of those joints would be zero.

[Vhunon]

I couldn't figure out how to set these 2 keypoints to 0 loss.
But I think its this line?
Also here:
loss = tf.reduce_mean(tf.square(heatmap_outs - gt_heatmap) * valid_mask)

[mks0601]

I think so. The valid mask is generated from joints[:,2]. The mask values would be zero for not annotated joints.