RGB camera (no events) evaluation

[samchopra2003]

Hi I have been able to run my own dataset using this codebase for events+rgb frames. I was wondering how I could run the codebase such that only my RGB frames are processed, and no events. Below is the config file I have prepared. While I have been able to run the code using only RGB frames, its resulting in a very low PSNR (around 0 and not improving):

num_gpu = 1
expname = falcon_indoor

basedir = logs
datadir = datasets/ev-deblurnerf_falcon_indoor/
tbdir = tb
dataset_type = llff

seed = 10000
factor = 1
llffhold = 5
llffhold_end

pts0_target_weight = 1.0
pts0_target_end_iter = 9999999

blur_loss_after = 1200
kernel_start_iter = 1200

pts0_target_start_iter = 0
add_event_egm_startiter = 0

tone_mapping_events_add_bii='none'

pose_transform_allknown = True
event_egm_use_colorevents = False
tone_mapping_learn_init_identity = True
tone_mapping_start_learn_iter = 1200
event_egm_use_color_weights = [0.4, 0.2, 0.4]
event_egm_color_weights_start_iter = 1200

events_N_rand 2048
events_tms_unit = 'us'
events_tms_files_unit = 'us'

use_events=False
events_threshold = 0.25

event_accumulate_step_range = [0, 0]
event_accumulate_step_range_end = [0, 0]
event_accumulate_step_scheduler = constant

event_egm_use_awp=False

add_event_egm=False
add_event_egm_stages = [stage0, stage1]
event_egm_weight = 0.1

N_rand = 1024
N_samples = 64
N_importance = 64
N_iters = 30000
lrate = 0.005
lrate_decay = 10

use_viewdirs = True
raw_noise_std = 1e0
rgb_activate = sigmoid

mode = c2f
coarse_num_layers = 2
coarse_num_layers_color = 3
coarse_hidden_dim = 64
coarse_hidden_dim_color = 64
coarse_app_dim = 32
coarse_app_n_comp = [64,16,16]
coarse_n_voxels = 16777248

fine_num_layers = 2
fine_num_layers_color = 3
fine_hidden_dim = 256
fine_hidden_dim_color = 256
fine_geo_feat_dim = 128
fine_app_dim = 32
fine_app_n_comp = [64,16,16]
fine_n_voxels = 134217984

kernel_type = RBK
kernel_ptnum = 10

kernel_rbk_use_viewdirs
kernel_img_embed = 32
kernel_rand_embed = 2  # the in_embed
kernel_spatial_embed = 2
kernel_depth_embed = 0

kernel_rbk_use_origin

kernel_rbk_extra_feat_ch = 0
kernel_rbk_se_r_depth = 1
kernel_rbk_se_r_width = 32
kernel_rbk_se_r_output_ch = 3
kernel_rbk_se_v_depth = 1
kernel_rbk_se_v_width = 32
kernel_rbk_se_v_output_ch = 3
kernel_rbk_ccw_depth = 1
kernel_rbk_ccw_width = 32
kernel_rbk_se_rv_window = 0.1

kernel_use_awp
kernel_awp_use_coarse_to_fine_opt
kernel_awp_sam_emb_depth=4
kernel_awp_sam_emb_width=64
kernel_awp_mot_emb_depth=1
kernel_awp_mot_emb_width=32
kernel_awp_dir_freq=2
kernel_awp_rgb_freq=2
kernel_awp_depth_freq=3
kernel_awp_ray_dir_freq=2

tone_mapping_type none
tone_mapping_events_type none

render_radius_scale = 0.5

I have removed use_pts0_prior = edi line from the config, as we dont need the prior from the event double integral for supervision.

[samchopra2003]

If I don't remove use_pts0_prior = edi from the config file then I get this error:

[samchopra2003]

Oh I realized that making these modifications to the config file allows me to train using only rgb frames properly. Looks like I was not training using the rgb frames generated loss (i.e blur loss) before:

blur_loss_after = 0
kernel_start_iter = 0