Test PSNR is higher than train PSNR.

[bo233]

trafficstars

Hi, there. Thank you for your awesome work! But when I study on your work, I meet something strange that the test PSNR is higher than train PSNR, as the following image shows.

Specifically, the result at 30k step is:

• test/psnr: 37.73
• test/psnr_all: 36.51
• train/psnr: 35.51

I check the PSNR calculation code, but everything is correct. I don't know where the problem is. I wonder whether you met this situation before.

Additionally, I train on the lego scene with the following config:

n_iters = 30000
batch_size = 4096

N_voxel_init = 2097156 # 128**3
N_voxel_final = 27000000 # 300**3
upsamp_list = [2000,3000,4000,5500,7000]
update_AlphaMask_list = [2000,4000]

N_vis = 5
vis_every = 10000

render_test = 1

n_lamb_sigma = [16,16,16]
n_lamb_sh = [48,48,48]
model_name = TensorVMSplit


shadingMode = MLP_Fea
fea2denseAct = softplus

view_pe = 2
fea_pe = 2

L1_weight_inital = 8e-5
L1_weight_rest = 4e-5
rm_weight_mask_thre = 1e-4

## please uncomment following configuration if hope to training on cp model
#model_name = TensorCP
#n_lamb_sigma = [96]
#n_lamb_sh = [288]
#N_voxel_final = 125000000 # 500**3
#L1_weight_inital = 1e-5
#L1_weight_rest = 1e-5