Test script.

[chenkang455]

Hi @LonsonZheng , thanks for your nice work. Will you kindly tell me how to write the code for evaluating RSIR on the real-world data (from .dat directly)? seq, label, length = load_spike_numpy(cfg.simulated_dir+"spike-video{}-00000-light{}.npz".format(tag, ls)), where can i get the label for the real-world data.

[LonsonZheng]

Hi @LonsonZheng , thanks for your nice work. Will you kindly tell me how to write the code for evaluating RSIR on the real-world data (from .dat directly)? seq, label, length = load_spike_numpy(cfg.simulated_dir+"spike-video{}-00000-light{}.npz".format(tag, ls)), where can i get the label for the real-world data.

We don't have the label for the real-world data, and the tests on real scenarios in the paper are calculated using non-reference image quality metrics.

[chenkang455]

Hi @LonsonZheng , thanks for your nice work. Will you kindly tell me how to write the code for evaluating RSIR on the real-world data (from .dat directly)? seq, label, length = load_spike_numpy(cfg.simulated_dir+"spike-video{}-00000-light{}.npz".format(tag, ls)), where can i get the label for the real-world data.

We don't have the label for the real-world data, and the tests on real scenarios in the paper are calculated using non-reference image quality metrics.

@LonsonZheng , thanks for your response. Bwt, i wonder if my script is right or not:

      Q, Nd, Nl = cal_para()
      q = torch.from_numpy(Q).cuda()
      nd = torch.from_numpy(Nd).cuda()
      nl = torch.from_numpy(Nl).cuda()
      for i in range(10):
          noisy_img = torch.ones([1, 1, spike.shape[2], spike.shape[3]], dtype=torch.float32).cuda()
          noisy_img[0, 0, :, :] = spike[:,16*i:16*(i+1)].mean(dim=1).float()
          if i == 0:
              input = noisy_img
          else:
              input = torch.cat([noisy_img, fusion_out], dim=1)
          fpn_denoise, img_true, fusion_out, denoise_out, refine_out, ft_denoise_out_d0, fgt_d0 = spkrecon_net(input, noisy_img, q, nd, nl)
          recon_img = fusion_out

recon_img is reconstructed after 10 iterations, i.e., 160 spike frames correspond to 1 output.

[LonsonZheng]

Hi @LonsonZheng , thanks for your nice work. Will you kindly tell me how to write the code for evaluating RSIR on the real-world data (from .dat directly)? seq, label, length = load_spike_numpy(cfg.simulated_dir+"spike-video{}-00000-light{}.npz".format(tag, ls)), where can i get the label for the real-world data.

We don't have the label for the real-world data, and the tests on real scenarios in the paper are calculated using non-reference image quality metrics.

@LonsonZheng , thanks for your response. Bwt, i wonder if my script is right or not:

      Q, Nd, Nl = cal_para()
      q = torch.from_numpy(Q).cuda()
      nd = torch.from_numpy(Nd).cuda()
      nl = torch.from_numpy(Nl).cuda()
      for i in range(10):
          noisy_img = torch.ones([1, 1, spike.shape[2], spike.shape[3]], dtype=torch.float32).cuda()
          noisy_img[0, 0, :, :] = spike[:,16*i:16*(i+1)].mean(dim=1).float()
          if i == 0:
              input = noisy_img
          else:
              input = torch.cat([noisy_img, fusion_out], dim=1)
          fpn_denoise, img_true, fusion_out, denoise_out, refine_out, ft_denoise_out_d0, fgt_d0 = spkrecon_net(input, noisy_img, q, nd, nl)
          recon_img = fusion_out

recon_img is reconstructed after 10 iterations, i.e., 160 spike frames correspond to 1 output.

Yes, this snippet of code appears to be accurate.

[chenkang455]

Thanks a lot for your feedback !