PointPillars inference pipeline (#153)

* Fix missing argument * more import bugfixes (#126) * fix missing attribute 'utils' * removed import statements for unused modules * fix conflict * add kitti dataset * add readme about config files (#129) * readme about config files * link to predefined script * Update README.md * address reviews * Simplified install instructions (#127) * Simpliefied install instructions Multiple requirements files for different use cases. * minor change to README * minor change adding link to docs for build from source. * update install instruction, add +cpu to torch requirements.txt Co-authored-by: Benjamin Ummenhofer <[email protected]> * added conditionals in requirements-torch for macos (#131) * Fix color dictionary in Semantic-KITTI bgr -> rgb (#130) * Enable hierarchy UI feature in visualizer for SemanticKITTI (#133) * PointPillars port from mmdet without NMS and Voxelizer. Only inference. Focal loss and smooth L1 implemented and tested. * loss function implemented and tested * Add bounding boxes to visualizer * Voxelization layer for point_pillars * apply style * Dataset bounding boxes are now fully their own object and do hide auto-hide/show pieces associated with a geometry, but they do auto-hide irrelevant boxes when animating. * Style fixes * objdet metric * removed mmdet3d dependency, o3d nms and voxelizer added * renamed objdet to run_pipeline, implemented part of the objdet pipeline * Added documentation, changed definition of size argument to be edge-to-edge and adjusted code accordingly * kitti eval only 3d boxes * test pipeline point pillars * howto small fix * change bbox class * improve bbox * style * visualization of results * test vis of results * using new bounding box class * waymo bbox * added nuscenes bbox * added lyft bbox * apply style * anchor head refactoring * fix waymo bbox * fix kitti bbox * apply style; * removed transformation of the predicted boxes * added licenses to files * removed unnecessary parts * format * Labels displayed in visualization. Kitti dataset label ordering changed to match MMDetection3D. * using our checkpoint format * fix ci * apply style * apply style * changed list to tuple Co-authored-by: Sanskar Agrawal <[email protected]> Co-authored-by: Benjamin Ummenhofer <[email protected]> Co-authored-by: YilingQiao <[email protected]> Co-authored-by: ssheorey <[email protected]> Co-authored-by: Ignacio Vizzo <[email protected]> Co-authored-by: prewettg <[email protected]> Co-authored-by: lprantl <[email protected]>
isl-org · Dec 8, 2020 · f76c093 · f76c093
1 parent 4862fb7
commit f76c093
Show file tree

Hide file tree

Showing 25 changed files with 1,556 additions and 21 deletions.
diff --git a/docs/howtos.md b/docs/howtos.md
@@ -54,7 +54,7 @@ Second, we will construct the networks and pipelines, load the pretrained weight
     pcs = get_custom_data(pc_names, data_path)
 ```
 
-Third, we can run the inference and collect the results and send the results to `Visualizer.visualize(list_of_pointclouds_to_visualize)`. Note that the input to `visualize()` visualize is a list of point clouds and their predictions. Each point cloud is a dictionary like, 
+Third, we can run the inference and collect the results and send the results to `Visualizer.visualize(list_of_pointclouds_to_visualize)`. Note that the input to `visualize()` is a list of point clouds and their predictions. Each point cloud is a dictionary like, 
 ```python
     vis_d = {
         "name": name,

diff --git a/ml3d/configs/pointpillars_kitti.yml b/ml3d/configs/pointpillars_kitti.yml
@@ -0,0 +1,43 @@
+dataset:
+  name: KITTI
+  dataset_path: # path/to/your/dataset
+  cache_dir: ./logs/cache
+
+model:
+  name: PointPillars
+  ckpt_path: # path/to/your/checkpoint
+
+  voxel_size: [0.16, 0.16, 4]
+  point_cloud_range: [0, -39.68, -3, 69.12, 39.68, 1]
+
+  voxelize:
+    max_num_points: 32
+    max_voxels: [16000, 40000]
+
+  voxel_encoder:
+    in_channels: 4
+    feat_channels: [64]
+
+  scatter:
+    in_channels: 64
+    output_shape: [496, 432]
+
+  backbone:
+    in_channels: 64
+    out_channels: [64, 128, 256]
+    layer_nums: [3, 5, 5]
+    layer_strides: [2, 2, 2]
+
+  neck:
+    in_channels: [64, 128, 256]
+    out_channels: [128, 128, 128]
+    upsample_strides: [1, 2, 4]
+    use_conv_for_no_stride: False
+
+  head:
+    num_classes: 3
+    in_channels: 384
+    feat_channels: 384
+
+pipeline:
+  name: ObjectDetection
diff --git a/ml3d/datasets/kitti.py b/ml3d/datasets/kitti.py
@@ -66,7 +66,13 @@ def __init__(self,
 
     @staticmethod
     def get_label_to_names():
-        label_to_names = {0: 'Car', 1: 'Pedestrian', 2: 'Cyclist', 3: 'Van'}
+        label_to_names = {
+            0: 'Pedestrian',
+            1: 'Cyclist',
+            2: 'Car',
+            3: 'Van',
+            4: 'DontCare'
+        }
         return label_to_names
 
     @staticmethod
@@ -172,7 +178,7 @@ def get_split_list(self, split):
         elif split in ['test', 'testing']:
             return self.test_files
         elif split in ['val', 'validation']:
-            return val_files
+            return self.val_files
         elif split in ['all']:
             return self.train_files + self.val_files + self.test_files
         else:
@@ -261,11 +267,11 @@ def cls_type_to_id(cls_type):
         get object id from name.
         """
         type_to_id = {
-            'DontCare': 0,
-            'Car': 1,
-            'Pedestrian': 2,
-            'Cyclist': 3,
-            'Van': 4
+            'Pedestrian': 0,
+            'Cyclist': 1,
+            'Car': 2,
+            'Van': 3,
+            'DontCare': 4,
         }
         if cls_type not in type_to_id.keys():
             return 0

diff --git a/ml3d/torch/models/__init__.py b/ml3d/torch/models/__init__.py
@@ -2,5 +2,6 @@
 
 from .randlanet import RandLANet
 from .kpconv import KPFCNN
+from .point_pillars import PointPillars
 
-__all__ = ['RandLANet', 'KPFCNN']
+__all__ = ['RandLANet', 'KPFCNN', 'PointPillars']