poor results on street scene

[LivingTom]

dear author, thanks for your open source, and this is a wonderful work.
I want to try my custom dataset, which is the street scene. And i have 615 images, all images were captured by the front-facing camera. And my training command is:

and use SDF to extract mesh, then i get result like:

I noticed that only the road area showed poor reconstruction results, while other areas performed better. Cloud you please provide some advice about improving the reconstruction effect of road areas. Hope for your reply~

[Jaykumaran]

Hi,

MIlo doesnt work well for large scenes. A practical workaround would be train 100-120 images and check if it improves the mesh quality

also, use radegs

python train.py -s seq3_curate_realityCapture/realit -m  /media/output/seq3_curate --rasterizer radegs --mesh_config veryhighres --data_device cpu --log_interval 200 --depth_order --depth_order_config boosted_gm --imp_metric outdoor 

python mesh_extract_sdf.py -s /output_0_105    -m output/V2_rcap_forward_0_105 --rasterizer radegs     --config veryhighres     --data_device cpu     --init learnable     --refine_iter 3000

depth_order/boosted_gm.yaml -- my custom config which gave good resultys

# Keep these for backwards compatibility or if you switch back to v2
depthanythingv2_checkpoint_dir: '../submodules/Depth-Anything-V2/checkpoints/'
depthanythingv2_encoder: 'vitl'

# --- THE TUNED SCHEDULE ---
# Default starts at 0.0, ramps to 0.1, then decays.
# This "Boosted" config ramps to 0.5 (Strong!) during the Geometric Phase (3k-8k).

weight_initial_value: 0.0

weight_update_iters:  [600, 3000, 8000, 15000]
weight_update_values: [0.1, 0.5,  0.2,   0.01]
# Not so good; weight_update_values: [0.9, 0.9,  0.09,   0.009]
# Explanation of schedule:
# Iter 0-600:   Weight 0.0 (Warmup, purely visual).
# Iter 600+:    Weight 0.1 (Gentle introduction, prevents NaN).
# Iter 3000+:   Weight 0.5 (High pressure! This flattens the road and sharpens edges before meshing).
# Iter 8000+:   Weight 0.2 (Lower pressure while the Mesh optimization takes over).

# --- SHARPNESS SETTINGS ---

# depth_ratio: How much we trust "Median" depth vs "Expected" depth.
# 1.0 = Pure Median (Sharper, ignores transparent fog).
# 0.6 = Mix (Default).
# Recommendation: Increase to 0.8 or 0.9 for sharper surfaces.
depth_ratio: 0.9

# max_pixel_shift_ratio: How far apart do we compare pixels?
# Large (0.05) = Good for global shape, bad for fine edges.
# Small (0.01) = Forces sharp edges between neighbors.
# Recommendation: Reduce to 0.02 to tighten object boundaries.
max_pixel_shift_ratio: 0.02

normalize_loss: true
log_space: true
log_scale: 20.0
reduction: 'mean'

[LivingTom]

Hi,

MIlo doesnt work well for large scenes. A practical workaround would be train 100-120 images and check if it improves the mesh quality

also, use radegs

python train.py -s seq3_curate_realityCapture/realit -m  /media/output/seq3_curate --rasterizer radegs --mesh_config veryhighres --data_device cpu --log_interval 200 --depth_order --depth_order_config boosted_gm --imp_metric outdoor 

python mesh_extract_sdf.py -s /output_0_105    -m output/V2_rcap_forward_0_105 --rasterizer radegs     --config veryhighres     --data_device cpu     --init learnable     --refine_iter 3000

depth_order/boosted_gm.yaml -- my custom config which gave good resultys

# Keep these for backwards compatibility or if you switch back to v2
depthanythingv2_checkpoint_dir: '../submodules/Depth-Anything-V2/checkpoints/'
depthanythingv2_encoder: 'vitl'

# --- THE TUNED SCHEDULE ---
# Default starts at 0.0, ramps to 0.1, then decays.
# This "Boosted" config ramps to 0.5 (Strong!) during the Geometric Phase (3k-8k).

weight_initial_value: 0.0

weight_update_iters:  [600, 3000, 8000, 15000]
weight_update_values: [0.1, 0.5,  0.2,   0.01]
# Not so good; weight_update_values: [0.9, 0.9,  0.09,   0.009]
# Explanation of schedule:
# Iter 0-600:   Weight 0.0 (Warmup, purely visual).
# Iter 600+:    Weight 0.1 (Gentle introduction, prevents NaN).
# Iter 3000+:   Weight 0.5 (High pressure! This flattens the road and sharpens edges before meshing).
# Iter 8000+:   Weight 0.2 (Lower pressure while the Mesh optimization takes over).

# --- SHARPNESS SETTINGS ---

# depth_ratio: How much we trust "Median" depth vs "Expected" depth.
# 1.0 = Pure Median (Sharper, ignores transparent fog).
# 0.6 = Mix (Default).
# Recommendation: Increase to 0.8 or 0.9 for sharper surfaces.
depth_ratio: 0.9

# max_pixel_shift_ratio: How far apart do we compare pixels?
# Large (0.05) = Good for global shape, bad for fine edges.
# Small (0.01) = Forces sharp edges between neighbors.
# Recommendation: Reduce to 0.02 to tighten object boundaries.
max_pixel_shift_ratio: 0.02

normalize_loss: true
log_space: true
log_scale: 20.0
reduction: 'mean'

thanks a lot !