Computer Vision News - October 2022

11 Intro to NeRF psnr = -10.0 * torch.log((( test_render - testimg) ** 2).mean()) / 2.30258509299 psnrs.append(psnr.cpu().detach().numpy()) iternums.append(iteration) plt.figure(figsize=(15,4)) plt.subplot(131) plt.title(f'Iteration: {iteration}') plt.imshow(test_render[0].detach().cpu().numpy()) plt.subplot(132) plt.title('Ground Truth') plt.imshow(testimg[0].detach().cpu().numpy()) plt.subplot(133) plt.plot(iternums, psnrs) plt.title('PSNR') plt.show() iteration += 1 After some epochs results look pretty good already. Different datasets are available online to play with NeRFs, and they include both synthetic and realistic data. NeRF can also be used to generate meshes, since they return a 3D shape explicitly as a continuous function. If you are interested in this application, don’t hesitate to let us know and we will talk more about it in the following months!