update readme

Author: Yue Pan

README.md

@@ -220,6 +220,10 @@ kiss_icp_pipeline <pc_path>
 The estimated pose file can be found in `./results/latest/velodyne.txt`. You can directly use it as your `pose_path`. In this case, you do not need a calib file, so just set `calib_path: ""` in the config file.
 </details>
 
+
+### Generate colorized mesh
+Check the repository [Color-SHINE-MAPPING](https://github.com/ZorAttC/COLOR_SHINE_MAPPING) (thanks [@ZorAttC](https://github.com/ZorAttC) for the contribution) for using SHINE Mapping to reconstruct colorized mesh using colorized point cloud.
+
 ----
 
 ## Run

config/config_icra2023/maicity_batch.yaml

@@ -20,11 +20,11 @@ process:
   vox_down_m: 0.05
   rand_down_r: 0.2
 sampler:
-  surface_sample_range_m: 0.3
+  surface_sample_range_m: 0.15
   surface_sample_n: 3
   free_sample_begin_ratio: 0.5
   free_sample_end_dist_m: 1.0
-  free_sample_n: 9
+  free_sample_n: 6
 octree:
   leaf_vox_size: 0.2
   tree_level_world: 12

config/kitti/kitti_batch.yaml

@@ -29,7 +29,7 @@ sampler:
 octree:
   leaf_vox_size: 0.3
   tree_level_world: 12
-  tree_level_feat: 2
+  tree_level_feat: 3
   feature_dim: 8
   poly_int_on: True
   octree_from_surface_samples: True

config/kitti/kitti_incre_replay.yaml

@@ -1,5 +1,6 @@
 setting:
   name: "kitti_incre_replay"
+  output_root: "./experiments/"
   pc_path: "xxx/data/kitti/sequences/00/velodyne"
   label_path: "xxx/data/kitti/sequences/00/labels" # optional for semantic mapping
   pose_path: "xxx/data/kitti/sequences/00/poses.txt"
@@ -14,19 +15,19 @@ setting:
   gpu_id: "0"
 process:
   min_range_m: 3.0
-  pc_radius_m: 25.0  # distance filter for each frame
+  pc_radius_m: 30.0  # distance filter for each frame
   min_z_m: -3.5
   rand_downsample: False # use random or voxel downsampling
-  vox_down_m: 0.05
+  vox_down_m: 0.08
   rand_down_r: 0.2
 sampler:
-  surface_sample_range_m: 0.5
+  surface_sample_range_m: 0.3
   surface_sample_n: 3
   free_sample_begin_ratio: 0.3
   free_sample_end_dist_m: 1.0
   free_sample_n: 3
 octree:
-  leaf_vox_size: 0.2
+  leaf_vox_size: 0.3
   tree_level_world: 12
   tree_level_feat: 3
   feature_dim: 8
@@ -50,17 +51,17 @@ continual:
   window_replay_on: True # replay within the sliding window
   window_radius_m: 50.0
 optimizer:
-  iters: 50  # iterations per frame
-  batch_size: 4096
+  iters: 100  # iterations per frame
+  batch_size: 8192
   learning_rate: 0.01
-  weight_decay: 1e-7 # l2 regularization
+  weight_decay: 0 # l2 regularization
 eval:
   wandb_vis_on: False # log to wandb or not
   o3d_vis_on: True # visualize the mapping or not
   vis_freq_iters: 0 
   save_freq_iters: 10000 # save the model and octree every x iterations
-  mesh_freq_frame: 10 # reconstruct the mesh every x frames
-  mc_res_m: 0.1 # reconstruction marching cubes resolution
+  mesh_freq_frame: 20 # reconstruct the mesh every x frames
+  mc_res_m: 0.15 # reconstruction marching cubes resolution
   mc_with_octree: True # querying sdf only in certain levels of the octree
   mc_local: False
   mc_vis_level: 1

config/maicity/maicity_batch.yaml

@@ -20,7 +20,7 @@ process:
   vox_down_m: 0.05
   rand_down_r: 0.2
 sampler:
-  surface_sample_range_m: 0.3
+  surface_sample_range_m: 0.15
   surface_sample_n: 3
   free_sample_begin_ratio: 0.3
   free_sample_end_dist_m: 0.8

config/maicity/maicity_incre_replay.yaml

@@ -14,19 +14,19 @@ setting:
   gpu_id: "0"
 process:
   min_range_m: 1.5
-  pc_radius_m: 25.0  # distance filter for each frame
+  pc_radius_m: 50.0  # distance filter for each frame
   min_z_m: -10.0
   rand_downsample: False # use random or voxel downsampling
-  vox_down_m: 0.03
+  vox_down_m: 0.05
   rand_down_r: 0.2
 sampler:
-  surface_sample_range_m: 0.5
+  surface_sample_range_m: 0.15
   surface_sample_n: 3
   free_sample_begin_ratio: 0.3
   free_sample_end_dist_m: 0.8
   free_sample_n: 3
 octree:
-  leaf_vox_size: 0.5
+  leaf_vox_size: 0.2
   tree_level_world: 12
   tree_level_feat: 3
   feature_dim: 8
@@ -42,16 +42,16 @@ loss:
   sigma_sigmoid_m: 0.05
   loss_weight_on: False
   behind_dropoff_on: False
-  ekional_loss_on: False
+  ekional_loss_on: True
   weight_e: 0.1
 continual:
   continual_learning_reg: False # using incremental mapping with regularization
   lambda_forget: 0 
   window_replay_on: True # replay within the sliding window
-  window_radius_m: 40.0
+  window_radius_m: 55.0
 optimizer:
   iters: 50  # iterations per frame
-  batch_size: 4096
+  batch_size: 8192
   learning_rate: 0.01
   weight_decay: 0 # l2 regularization
 eval:

config/ncd/ncd_incre_replay.yaml

@@ -17,7 +17,7 @@ process:
   pc_radius_m: 50.0  # distance filter for each frame
   min_z_m: -10.0
   rand_downsample: False # use random or voxel downsampling
-  vox_down_m: 0.05
+  vox_down_m: 0.1
   rand_down_r: 0.2
 sampler:
   surface_sample_range_m: 0.5
@@ -51,7 +51,7 @@ continual:
   window_radius_m: 60.0
 optimizer:
   iters: 50  # iterations per frame
-  batch_size: 4096
+  batch_size: 8192
   learning_rate: 0.01
   weight_decay: 1e-7 # l2 regularization
 eval:

config/rgbd/rgbd_incre_replay.yaml

@@ -17,7 +17,7 @@ process:
   pc_radius_m: 5.0  # distance filter for each frame
   min_z_m: -10.0 
   rand_downsample: False # use random or voxel downsampling
-  vox_down_m: 0.01
+  vox_down_m: 0.02
   rand_down_r: 0.2
 sampler:
   surface_sample_range_m: 0.05
@@ -48,7 +48,7 @@ continual:
   continual_learning_reg: False # using incremental mapping with regularization
   lambda_forget: 0 
   window_replay_on: True # replay within the sliding window
-  window_radius_m: 6.0
+  window_radius_m: 7.0
 optimizer:
   iters: 50
   batch_size: 4096

dataset/lidar_dataset.py

@@ -91,7 +91,7 @@ def __init__(self, config: SHINEConfig, octree: FeatureOctree = None) -> None:
         # or we directly use the world frame as reference
 
         # to cope with the gpu memory issue (use cpu memory for the data pool, a bit slower for moving between cpu and gpu)
-        if self.used_pc_count > config.pc_count_gpu_limit and not config.continual_learning_reg:
+        if self.used_pc_count > config.pc_count_gpu_limit and not config.continual_learning_reg and not config.window_replay_on:
             self.pool_device = "cpu"
             self.to_cpu = True
             self.sampler.dev = "cpu"