Fail to reproduce performance in some scenes

[bbangsik13]

Dear Authors,

Thank you for sharing your research.

I have some issue when running your code on the n3v dataset, where some scenes produce unexpected results.
I followed the dataset preparation steps following the readme file and run the code as follow:

for SCENE in {coffee_martini_wo_cam13,cook_spinach,cut_roasted_beef,flame_salmon_frag1,flame_steak,sear_steak}
do
python opt.py -t pretrained/${SCENE} data/${SCENE}/frames/0000 \
 -c configs/n3dv_init.json --offset 500 --scale 0.5 --nosphereinit
python train_video_n3dv_pilot.py -t outputs/${SCENE} data/${SCENE}/frames \
 -c configs/n3dv.json --batch_size 20000 --pretrained pretrained/${SCENE}/ckpt.npz --n_iters 750 --lr_sigma 1.0 \
 --lr_sigma_final 1.0  --lr_sh 1e-2 --lr_sh_final 1e-3 --lr_sigma_decay_steps 750 --lr_sh_decay_steps 750 \
 --frame_end 300 --fps 30 --train_use_all 0 --offset 750  --scale 0.5 --sh_keep_thres 0.5 --sh_prune_thres 0.1 \
 --performance_mode  --dilate_rate_before 1 --dilate_rate_after 1 --stop_thres 0.01  --compress_saving \
 --save_delta --pilot_factor 2
python train_video_n3dv_full.py -t outputs/${SCENE} data/${SCENE}/frames \
 -c configs/n3dv_full.json --batch_size 20000 --pretrained pretrained/${SCENE}/ckpt.npz --n_iters 500 \
 --lr_sigma 1.0 --lr_sigma_final 1.0  --lr_sh 1e-2 --lr_sh_final 3e-3 --lr_sigma_decay_steps 500 \
 --lr_sh_decay_steps 300   --frame_end 300 --fps 30 --train_use_all 0 --offset 1500  --scale 0.5 --sh_keep_thres 1.0 \
 --sh_prune_thres 0.2 --performance_mode  --dilate_rate_before 2 --dilate_rate_after 2  --stop_thres 0.01  \
 --compress_saving --save_delta  --apply_narrow_band
done

While the pilot results looks good, the full_model training fails as follow:

coffee_martini_wo_cam13_pilot.mp4
coffee_martini_wo_cam13_full.mp4

I also included the full_model logs below.
log_full.log

Please let me know if there is a mistake.
Thank you for your assistance.