add example configs

Author: striest

config/image_processing/example.yaml

@@ -0,0 +1,14 @@
+# example config for image processing pipeline
+# expected format should be a list of blocks (defined in physics_atv_visual_mapping/image_processing/image_pipeline.py)
+# this will define how to extract "features" from the input set of images
+-
+    type: radio_lang
+    args:
+        image_insize: [848, 448]
+        radio_type: c-radio_v3-b
+        adaptor_type: siglip2
+
+- 
+    type: pca
+    args:
+        fp: physics_atv_visual_mapping/pca/radio_v3b_siglip2.pt

config/localmapping/example.yaml

@@ -0,0 +1,15 @@
+
+## example localmapping config
+
+mapper_type: voxel                      ## one of {bev, voxel}. Whether to map in 2.5d or 3d
+ema: 0.5                                ## how to merge multiple features together (output will be ema*new + (1-ema)*old)
+n_features: -1                          ## number of features to map per cell. if -1, matches the number of input features
+metadata:                               ## metadata of the space to map (2d if bev, 3d if voxel)
+    origin: [-50., -50., -10.]          ## position (m) of lower-bottom-left corner of map w.r.t. pose
+    length: [100., 100., 20.]           ## position (m) of upper-top-right corner w.r.t. origin
+    resolution: [0.4, 0.4, 0.1]         ## size (m) of a cell
+
+raytracer:                              ## optional raytracing config for voxels
+    type: frustum
+    sensor:
+        type: VLP32C-front

config/ros/yamaha_voxel_mapping.yaml

@@ -1,9 +1,13 @@
+## define location of models e.g. Dino/RADIO/etc.
 models_dir: /home/tartandriver/tartandriver_ws/models #necessary to include this for non-ROS offline proc
 
+## define image processing pipeline (e.g. how to compute image-space features from raw)
 image_processing: !include /home/tartandriver/tartandriver_ws/src/perception/physics_atv_visual_mapping/config/image_processing/radiov3_siglip2_pca.yaml
 
+## define localmapping representation (e.g. voxel/BEV, mapping extent, etc.)
 localmapping: !include /home/tartandriver/tartandriver_ws/src/perception/physics_atv_visual_mapping/config/localmapping/voxel.yaml
 
+## optionally, define how to compute a terrain representation from mapping representation
 terrain_estimation: !include /home/tartandriver/tartandriver_ws/src/perception/physics_atv_visual_mapping/config/terrain_estimation/voxel_to_bev.yaml
 
 device: cuda

config/terrain_estimation/example..yaml

@@ -0,0 +1,108 @@
+# example terrain estimation config, i.e. converting from localmapping representation to 2.5d
+# note that we can perform terrain estimation on both BEV and voxel localmapping
+# but some blocks only work on BEV or voxel
+#input should be a list of blocks to run in sequence
+
+-
+    ## compute min/max elevation of voxel grid per cell
+    type: elevation_stats
+    args:
+        use_voxel_centers: false
+
+-
+    ## compute average porosity (hits/(hits+misses)) in a BEV column
+    type: porosity
+    args: {}
+
+-
+    ## filter out non-smooth changes in the elevation map
+    type: elevation_filter
+    args:
+        input_layer: min_elevation
+        cnt_layer: num_voxels
+
+        height_low_thresh: -3.0 #cells this far below their neighbors are not terrain
+        height_high_thresh: 0.5 #cells this far above their neighbors are not terrain
+
+        kernel_params:
+            kernel_type: gaussian #one of {gaussian/box} the kernel type to use for inflation
+
+            kernel_radius: 2. #kernel radius in m
+            kernel_sharpness: 5. #sharpness of (Gaussian) kernel
+
+            # kernel_radius: 0.25 #kernel radius in m
+            # kernel_sharpness: 0.5 #sharpness of (Gaussian) kernel
+
+-
+    ## fill in holes in elevation map with 2d conv
+    type: terrain_inflation
+    args:
+        input_layer: min_elevation_filtered
+        mask_layer: min_elevation_filtered_mask
+
+        thresh: 0.05 #at least this frac of neighboring cells in the kernel must be observed
+
+        kernel_params:
+            kernel_type: gaussian #one of {gaussian/box} the kernel type to use for inflation
+
+            kernel_radius: 5. #kernel radius in m
+            kernel_sharpness: 2. #sharpness of (Gaussian) kernel
+
+            # kernel_radius: 0.5 #kernel radius in m
+            # kernel_sharpness: 1.0 #sharpness of (Gaussian) kernel
+
+- 
+    ## refine terrain estimate using Markov Random Field
+    type: mrf_terrain_estimation
+    args: 
+        input_layer: min_elevation_filtered_inflated
+        mask_layer: min_elevation_filtered_inflated_mask
+
+        itrs: 5   #num updates
+        alpha: 10. #weight on the measurement update
+        beta: 1.  #weight on the neighbor update
+        lr: 0.05   #learning rate
+
+        kernel_params:
+            kernel_type: gaussian #one of {gaussian/box} the kernel type to use for inflation
+
+            kernel_radius: 1. #kernel radius in m
+            kernel_sharpness: 2. #sharpness of (Gaussian) kernel
+
+            # kernel_radius: 0.5
+            # kernel_sharpness: 1.0
+
+- 
+    ## compute slope as spatial derivatives of terrain
+    type: slope
+    args:
+        input_layer: terrain
+        mask_layer: min_elevation_filtered_inflated_mask
+        radius: 0.25
+        max_slope: 2.0
+        kernel_type: scharr
+
+-
+    ## compute terrain normal vectors from slope values
+    type: normals_sobel
+    args:
+        slope_x_layer: slope_x
+        slope_y_layer: slope_y
+        mask_layer: min_elevation_filtered_inflated_mask
+
+-
+    ## compute max height of voxels within overhang of terrain
+    type: terrain_diff
+    args:
+        terrain_layer: terrain
+        overhang: 2.0
+
+-
+    ## splat voxels within overhang of terrain into BEV
+    type: terrain_aware_bev_feature_splat
+    args:
+        metainfo_key: vfm
+        n_features: 16
+        terrain_layer: terrain
+        terrain_mask_layer: min_elevation_filtered_inflated_mask
+        overhang: 2.0