add monoscene lite

.gitignore

@@ -1,4 +1,4 @@
-__pycache__
+*__pycache__*
 .ipynb_checkpoints
 *.ckpt
 gradio*

app.py

@@ -10,16 +10,13 @@ from monoscene.monoscene import MonoScene
 csv.field_size_limit(sys.maxsize)
 torch.set_grad_enabled(False)
 
-# pipeline = pipeline(model="anhquancao/monoscene_kitti")
-# model = AutoModel.from_pretrained(
-#     "anhquancao/monoscene_kitti", trust_remote_code=True, revision='bf033f87c2a86b60903ab811b790a1532c1ae313'
-# )#.cuda()
+
 model = MonoScene.load_from_checkpoint(
         "monoscene_kitti.ckpt",
         dataset="kitti",
         n_classes=20,
         feature = 64,
-        project_scale = 2,
+        project_scale = 4,
         full_scene_size = (256, 256, 32),
     )
 
@@ -45,19 +42,16 @@ def predict(img):
         batch[k] = batch[k].unsqueeze(0)#.cuda()
 
     pred = model(batch).squeeze()
-    # print(pred.shape)
-    # pred = majority_pooling(pred, k_size=2)
-    fig = draw(pred, batch['fov_mask_1'])
+    fig = draw(pred, batch['fov_mask_2'])
 
 
     return fig
    
-# The output is <b>downsampled by 2</b> to be able to be rendered in browsers. 
 
 description = """
 MonoScene Demo on SemanticKITTI Validation Set (Sequence 08), which uses the <b>camera parameters of Sequence 08</b>.
 Due to the <b>CPU-only</b> inference, it might take up to 20s to predict a scene. \n
-<b>Darker</b> colors represent the <b>scenery outside the Field of View</b>, i.e. not visible on the image.
+This is a <b>smaller</b> model with half resolution and <b>w/o 3D CRP</b>. You can find the full model at: <a href="https://huggingface.co/spaces/CVPR/MonoScene">https://huggingface.co/spaces/CVPR/MonoScene</a>
 <center>
     <a href="https://cv-rits.github.io/MonoScene/">
         <img style="display:inline" alt="Project page" src="https://img.shields.io/badge/Project%20Page-MonoScene-red">
@@ -66,15 +60,17 @@ Due to the <b>CPU-only</b> inference, it might take up to 20s to predict a scene
     <a href="https://github.com/cv-rits/MonoScene"><img style="display:inline" src="https://img.shields.io/github/stars/cv-rits/MonoScene?style=social"></a>
 </center>
 """
-title = "MonoScene: Monocular 3D Semantic Scene Completion"
+title = "MonoScene Lite - Half resolution, w/o 3D CRP"
 article="""
 <center>
-    <img src='https://visitor-badge.glitch.me/badge?page_id=anhquancao.MonoScene&left_color=darkmagenta&right_color=purple' alt='visitor badge'>
+    <img src='https://visitor-badge.glitch.me/badge?page_id=anhquancao.MonoScene_lite&left_color=darkmagenta&right_color=purple' alt='visitor badge'>
 </center>
 """
 
 examples = [
-    'images/08/000010.jpg',
+    'images/08/001385.jpg',
+    'images/08/000295.jpg',
+    'images/08/002505.jpg',
     'images/08/000085.jpg',
     'images/08/000290.jpg',
     'images/08/000465.jpg',
@@ -83,10 +79,10 @@ examples = [
     'images/08/001380.jpg',
     'images/08/001530.jpg',
     'images/08/002360.jpg',
-    'images/08/002505.jpg',
     'images/08/004059.jpg',
     'images/08/003149.jpg',
     'images/08/001446.jpg',
+    'images/08/000010.jpg',
     'images/08/001122.jpg',
     'images/08/003533.jpg',
     'images/08/003365.jpg',
@@ -108,6 +104,7 @@ examples = [
 
 
 
+
 demo = gr.Interface(
     predict, 
     gr.Image(shape=(1220, 370)), 

helpers.py

@@ -126,7 +126,7 @@ def get_grid_coords(dims, resolution):
     return coords_grid
 
 def get_projections(img_W, img_H):
-    scale_3ds = [1, 2]
+    scale_3ds = [2, 4]
     data = {}
     for scale_3d in scale_3ds:
         scene_size = (51.2, 51.2, 6.4)
@@ -188,7 +188,7 @@ def draw(
     fov_mask,
     # img_size,
     # f,
-    voxel_size=0.2,
+    voxel_size=0.4,
     # d=7,  # 7m - determine the size of the mesh representing the camera
 ):
 

monoscene/monoscene.py

@@ -96,15 +96,15 @@ class MonoScene(pl.LightningModule):
                 if x3d is None:
                     x3d = self.projects[str(scale_2d)](
                         x_rgb["1_" + str(scale_2d)][i],
-                        torch.div(projected_pix, scale_2d, rounding_mode='floor'),
-                        # projected_pix // scale_2d,
+                        # torch.div(projected_pix, scale_2d, rounding_mode='floor'),
+                        projected_pix // scale_2d,
                         fov_mask,
                     )
                 else:
                     x3d += self.projects[str(scale_2d)](
                         x_rgb["1_" + str(scale_2d)][i],
-                        torch.div(projected_pix, scale_2d, rounding_mode='floor'),
-                        # projected_pix // scale_2d,
+                        # torch.div(projected_pix, scale_2d, rounding_mode='floor'),
+                        projected_pix // scale_2d,
                         fov_mask,
                     )
             x3ds.append(x3d)

monoscene_kitti.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f0d1324885166f17949bf2dcfc0ee1eb2d2aedd0f48e75b56bb2beb87c1ce3a
-size 1796467007
+oid sha256:82b3bff496ca832738dd184d4c9f0d555deb5d26f0d4c6ce916b8b0ec6feab62
+size 1612174535