After I was training the whole CASIA-B dataset. about gaitset

[prb2307]

I think you can use the two checkpoints given and load model from there. I also trained the whole dataset. I am not sure whether I am making sense. Pretty new to this ;)
you can refer this repo for extracting human silhouettes: url

[AbnerHqC]

Hi~~ I believe the model trained on the whole CASIA-B dataset would has a better performance than the given checkpoints (trained with 74 persons). But I think you might have to spare a few of data for validation. So that you can find out which checkpoint you got during training is better.

As for test on custom video, besides the preprocessing like segmentation and alignment, I think you might also need to fine-tune the model on your own data first. Since data in CASIA-B is not so 'natural'.

[LavenderMP]

What should i do next to apply it on the camera?

[prb2307]

@LavenderMP you have to write your custom code for that. 1.take video 2. through the image segmentation get the silhouettes.

[LavenderMP]

Thanks for reply on my issue.
I know this is stupid but after i train the model and i got 20gb those checker file what do i do with that.
And btw how to make images like dataset from videos.
Btw i got the visdeos walking already

[prb2307]

@AbnerHqC Thanks. What do you mean by not so "natural"? Could you please elaborate, please? I tried to test it on the silhouettes I got by using OpenCV, the results are very bad as expected. (as the silhouettes are not of good quality) Have you tested the code with silhouettes other than CASIA-B?

[AbnerHqC]

@prashant-bansod In casia-b, the videos were shot under a specific scene. People were walking with no other gestures. The pitch angle is fixed. However, in a natural scene, the circumstances could be much more complicated. People might have various gestures. For instance, people might be holding some things in their hands. Thus, I guess to get a satisfying result you might need to fine-tune the model.

[AbnerHqC]

@prashant-bansod Current public gait datasets are very limited. We tested the model on casia-b and oumvlp as we presented in our paper but haven't tested it on any real scenarios.

[nezihkasim]

And, seems, there is no cross dataset tests... The model trained with CASIA-B, probably, cannot give satisfactory result for OUMVLP dataset, vice versa. This makes the model dataset-specific model, which is unfortunately unpractical for real scenarios :(