A fine-tuning of GaitSet and evaluate it on CASIA-B dataset.
We add a new module to extract the fine-grained feature. And we also add a multi-layer feature fusion module to extract complementary information.
We evaluate our methods on CASIA-B dataset and compare it with GaitSet(AAAI 2018) and GaitPart(CVPR2020). Because of our lab's server's problem, we can't evalute the method on the largest dataset named OU-MVLP.
model | NM | BG | CL |
---|---|---|---|
GaitSet | 95.0% | 87.2% | 70.4% |
Ours(Fine-tune) | 96.8% | 91.8% | 78.0% |
model | NM | BG | CL |
---|---|---|---|
GaitPart | 96.2% | 91.5% | 78.7% |
Ours(Fine-tune) | 96.8% | 91.8% | 78.0% |
As we can see, We surpass GaitSet on three differnet sequences type. We also surpass GaitPart on NM and BG, but our CL's accuracy is lower than GaitPart.
Maybe there are tremendous tricks to boost the accuracy. Hope readers' commands!!!