/PoseEncoding

Reimplementation for paper pose encoding for robust skeleton-based action recognition

Primary LanguagePython

CVPRW复现

  • Result
Epoch 250
    default_norm
        Best Epoch 221, Test Acc 75.92%
    recovered
        Best Epoch 198, Test Acc 79.39%
    hidden
        Best Epoch 249, Test Acc 83.08%
Epoch 140
    default_norm
        Best Epoch 134, Test Acc 69.63%
    recovered
        Best Epoch 137, Test Acc 75.70%
    hidden
        Best Epoch 135, Test Acc 74.19%

  • Baseline

    sh ./scripts/default_norm.sh

  • Generate clean data from auto-encoder

    sh ./scripts/train_ae_trail.sh sh ./scripts/use_ae_trail.sh

  • Final result

    sh ./scripts/hidden.sh

    and

    sh ./scripts/recovered.sh

  • Visualization

    Raw data: python3 feeder.py --local --vid a01_s01_e00_v2_skeleton --no_norm

    Normed data: python3 feeder.py --local --vid a01_s01_e00_v2_skeleton --norm

    Final clean data: python3 feeder.py --local --vid a01_s01_e00_v2_skeleton --modality _recovered