a bug of ws-psnr calculation #6
Description
Hi! In your code, the input shape required to calculate ws-psnr is (B, H, W, C), and your input is (B, C, H, W), which causes the calculated ws-psnr to be a little higher than expected. Could you check it out? If I make any mistakes, please point them out! This is my revised inference result:
m3d_0.1_psnr 33.929344749450685
m3d_0.1_lpips 0.06606819555163383
m3d_0.1_ssim 0.9587008893489838
m3d_0.1_ws_psnr 32.879531669616696
m3d_0.25_psnr 32.49078884124756
m3d_0.25_lpips 0.07195058465003967
m3d_0.25_ssim 0.9504191100597381
m3d_0.25_ws_psnr 31.448466491699218
abs_rel 0.01715618809685111
sq_rel 0.007986347749829292
rmse 0.14329498074948788
m3d_0.5_psnr 29.933135414123534
m3d_0.5_lpips 0.09054138734936715
m3d_0.5_ssim 0.9305436789989472
m3d_0.5_ws_psnr 28.805792808532715
m3d_0.75_psnr 24.75487403869629
m3d_0.75_lpips 0.18016190454363823
m3d_0.75_ssim 0.8492980301380157
m3d_0.75_ws_psnr 24.088992500305174
m3d_1.0_psnr 20.747944641113282
m3d_1.0_lpips 0.27218877971172334
m3d_1.0_ssim 0.7731463372707367
m3d_1.0_ws_psnr 20.194300174713135
residential_0.15_psnr 31.481586456298828
residential_0.15_lpips 0.17258246739705405
residential_0.15_ssim 0.9166631102561951
residential_0.15_ws_psnr 30.996205012003582
replica_0.5_psnr 31.650250964694553
replica_0.5_lpips 0.0695569028870927
replica_0.5_ssim 0.9612669944763184
replica_0.5_ws_psnr 30.74579069349501