Merge pull request #165 from princeton-vl/develop

Bugfix v1.4.1

docs/CHANGELOG.md

@@ -56,3 +56,8 @@ v1.4.0 - Infinigen Indoors
 - Add library of procedural generators for indoor objects & materials
 - Add indoor scene generation system, including constraint language and solver
 - Add HelloRoom.md & ExportingToSimulators.md
+
+v1.4.1
+- @David-Yan1 fix placeholder & ocmesher submodule version
+- @lahavlipson fix bug in surface normals of wall meshes
+- @araistrick bugfix example commands & other typos

docs/ConfiguringInfinigen.md

@@ -172,7 +172,7 @@ We recommend this command as a starting point for generating high quality videos
 ````
 python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_scenes 500 \
     --pipeline_config slurm monocular_video cuda_terrain opengl_gt \
-    --cleanup big_files --warmup_sec 60000 --config video high_quality_terrain
+    --cleanup big_files --warmup_sec 60000 --config trailer_video high_quality_terrain
 ````
 
 #### Creating large-scale stereo datasets
@@ -218,7 +218,7 @@ python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_
 ```
 python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_scenes 500 \
     --pipeline_config slurm monocular_video cuda_terrain opengl_gt \
-    --cleanup big_files --warmup_sec 30000 --config video high_quality_terrain \
+    --cleanup big_files --warmup_sec 30000 --config trailer_video high_quality_terrain \
     --overrides camera.camera_pose_proposal.altitude=["uniform", 20, 30]
 ```
 
@@ -228,7 +228,7 @@ python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_
 ```
 python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_scenes 500 \
     --pipeline_config slurm monocular_video cuda_terrain opengl_gt \
-    --cleanup big_files --warmup_sec 30000 --config video high_quality_terrain \
+    --cleanup big_files --warmup_sec 30000 --config trailer_video high_quality_terrain \
     --pipeline_overrides iterate_scene_tasks.frame_range=[1,25]
 ```
 

docs/HelloRoom.md

@@ -76,12 +76,12 @@ We also provide an OpenGL-based ground truth extractor which offers additional g
 
 To generate a single scene in one command, you can run the following:
 ```bash
-screen python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1000 --pipeline_configs local_256.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True 
+screen python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1000 --pipeline_configs local_256GB.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True 
 ```
 
 To create a large dataset of many random rooms, we recommend:
 ```bash
-screen python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1000 --pipeline_configs local_256.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True 
+screen python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1000 --pipeline_configs local_256GB.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True 
 ```
 
 You can inspect `outputs/my_dataset/SEED/` to see the running logs of the subprocesses and output results.

docs/Installation.md

@@ -98,7 +98,7 @@ INFINIGEN_MINIMAL_INSTALL=True bash scripts/install/interactive_blender.sh
 bash scripts/install/interactive_blender.sh
 
 # Enable OpenGL GT
-INFINIGEN_INSTALL_CUSTOMGT=True scripts/install/interactive_blender.sh
+INFINIGEN_INSTALL_CUSTOMGT=True bash scripts/install/interactive_blender.sh
 ```
 
 :exclamation: If you encounter any issues with the above, please add ` > logs.txt 2>&1` to the end of your command and run again, then provide the resulting logs.txt file as an attachment when making a Github Issue.

infinigen/__init__.py

@@ -1,3 +1,3 @@
 import logging
 
-__version__ = "1.4.0"
+__version__ = "1.4.1"

infinigen/assets/shelves/kitchen_cabinet.py

@@ -325,4 +325,4 @@ def sample_params(self):
 
     def create_placeholder(self, **kwargs) -> bpy.types.Object:
         x,y,z = self.dimensions
-        return new_bbox(-x/2 * 1.2, x/2 * 1.2, 0, y * 1.1, 0, (z + 0.06) * 1.03)
+        return new_bbox(-x/2 * 1.2, x/2 * 1.2, 0, y * 1.1, 0, (z + 0.06))

infinigen/assets/shelves/kitchen_space.py

@@ -145,7 +145,7 @@ def sample_parameters(self, dimensions):
 
     def create_placeholder(self, **kwargs) -> bpy.types.Object:
         x, y, z = self.dimensions
-        box = new_bbox(-x/2 * 1.08, x/2 * 1.08, 0, y, 0, self.cabinet_bottom_height + 0.13)
+        box = new_bbox(-x/2 * 1.08, x/2 * 1.08, 0, y, 0, self.cabinet_bottom_height + 0.095)
         surface.add_geomod(box, nodegroup_tag_cube, apply=True)
 
         if not self.island:

infinigen/core/execute_tasks.py

@@ -81,7 +81,7 @@
     pipeline, 
     exporting
 )
-from infinigen.tools.export import export_scene
+from infinigen.tools.export import export_scene, triangulate_meshes
 from infinigen.core.util.math import FixedSeed, int_hash
 from infinigen.core.util.logging import Timer, save_polycounts, create_text_file
 from infinigen.core.util.pipeline import RandomStageExecutor
@@ -228,29 +228,14 @@ def render(scene_seed, output_folder, camera_id, render_image_func=render_image,
     with Timer('Render Frames'):
         render_image_func(frames_folder=Path(output_folder), camera_id=camera_id)
 
-def triangulate_meshes():
-    for obj in bpy.context.scene.objects:
-        if obj.type == 'MESH':
-            view_state = obj.hide_viewport
-            obj.hide_viewport = False
-            bpy.context.view_layer.objects.active = obj
-            obj.select_set(True)
-            bpy.ops.object.mode_set(mode='EDIT')
-            bpy.ops.mesh.select_all(action='SELECT')
-            logging.info(f"Triangulating {obj}")
-            bpy.ops.mesh.quads_convert_to_tris()
-            bpy.ops.object.mode_set(mode='OBJECT')
-            obj.select_set(False)
-            obj.hide_viewport = view_state
-
 @gin.configurable
 def save_meshes(scene_seed, output_folder, frame_range, resample_idx=False):
 
     if resample_idx is not None and resample_idx > 0:
         resample_scene(int_hash((scene_seed, resample_idx)))
-    
+
     triangulate_meshes()
-    
+
     for obj in bpy.data.objects:
         obj.hide_viewport = obj.hide_render
 

infinigen/core/placement/camera.py

@@ -667,7 +667,7 @@ def save_camera_parameters(camera_ids, output_folder, frame, use_dof=False):
             bpy.context.scene.render.resolution_y, 
             bpy.context.scene.render.resolution_x
         ))
-        T = np.asarray(camera_obj.matrix_world, dtype=np.float64) @ np.diag((1.,-1.,-1.,1.))
+        T = np.asarray(camera_obj.matrix_world, dtype=np.float64) @ np.diag((1.,-1.,-1.,1.)) # Y down Z forward (aka opencv)
         np.savez(output_file, K=np.asarray(K, dtype=np.float64), T=T, HW=height_width)
 
 if __name__ == "__main__":

infinigen/datagen/customgt/camera_view.cpp

@@ -51,34 +51,31 @@ const Matrix4f FLIP_Y_Z = Eigen::Vector4f({1,-1,-1,1}).asDiagonal();
 
 template <int h, int w, typename T_orig=double, typename T_final=float>
 Eigen::Matrix<T_final, h, w> load_matrix(const npz &camview, const std::string &key){
-    const auto blender_camera_pose_data = camview.read_data<T_orig>(key);
-    const auto tmp = Eigen::Matrix<T_orig, h, w>(blender_camera_pose_data.data());
+    const auto opengl_camera_pose_data = camview.read_data<T_orig>(key);
+    const auto tmp = Eigen::Matrix<T_orig, h, w>(opengl_camera_pose_data.data());
     return tmp.transpose().template cast<T_final>();
 }
 
 CameraView::CameraView(const std::string fstr, const fs::path input_dir, const int width, const int height) : frame_string(fstr), image_height(height), image_width(width), buffer_height(height*2), buffer_width(width*2)
 {
     // Current Frame
     const fs::path current_frame_cam_path = input_dir / ("camview_"+frame_string+".npz");
-    const npz current_camview(current_frame_cam_path);
-    const Matrix4f blender_camera_pose = load_matrix<4, 4>(current_camview, "T") * FLIP_Y_Z; // TODO REMOVE
-    current_frame_view_matrix = glm::make_mat4(Matrix4f(blender_camera_pose.inverse()).data());
+    const npz current_camview(current_frame_cam_path); // Y down Z forward (aka opencv)
+    const Matrix4f opengl_camera_pose = load_matrix<4, 4>(current_camview, "T") * FLIP_Y_Z;
+    current_frame_view_matrix = glm::make_mat4(Matrix4f(opengl_camera_pose.inverse()).data()); // Y up Z back (aka blender/opengl)
 
     // Next Frame
     const fs::path next_frame_cam_path = increment_int_substr({"frame_([0-9]{4})", "camview_[0-9]+_[0-9]+_([0-9]{4})"}, current_frame_cam_path);
     const npz next_camview(next_frame_cam_path);
-    const Matrix4f next_blender_camera_pose = load_matrix<4, 4>(next_camview, "T") * FLIP_Y_Z; // TODO REMOVE
-    next_frame_view_matrix = glm::make_mat4(Matrix4f(next_blender_camera_pose.inverse()).data());
+    const Matrix4f next_opengl_camera_pose = load_matrix<4, 4>(next_camview, "T") * FLIP_Y_Z;
+    next_frame_view_matrix = glm::make_mat4(Matrix4f(next_opengl_camera_pose.inverse()).data());
 
-    // Set Camera Position
-    position = glm::make_vec3(blender_camera_pose.block<3, 1>(0, 3).data());
-
-    // Set WC -> Img Transformation
+    // Set CC -> Img Transformation
     const Matrix3f K_mat3x3 = load_matrix<3, 3>(current_camview, "K");
     Matrix4f K_mat = Matrix4f::Identity();
     buffer_over_image = 2;
     K_mat.block<2,3>(0, 0) = buffer_over_image * K_mat3x3.block<2,3>(0, 0);
-    wc2img = glm::make_mat4(Matrix4f(K_mat * FLIP_Y_Z * blender_camera_pose.inverse()).data());
+    cc2img = glm::make_mat4(K_mat.data()); // maps opencv wc to img
 
     fx = K_mat(0,0);
     fy = K_mat(1,1);
@@ -107,10 +104,9 @@ CameraView::CameraView(const std::string fstr, const fs::path input_dir, const i
 void CameraView::activateShader(Shader &shader) const {
     shader.use();
     shader.setMat4("projection", projection);
-    shader.setMat4("view", current_frame_view_matrix);
+    shader.setMat4("view", current_frame_view_matrix);// Y up Z back (aka blender/opengl)
     shader.setMat4("viewNext", next_frame_view_matrix);
-    shader.setMat4("wc2img", wc2img);
-    shader.setVec3("cameraPos", position);
+    shader.setMat4("cc2img", cc2img);
 }
 
 Tensor<double, 3> CameraView::project(const Tensor<double, 3> &cam_coords) const {

infinigen/datagen/customgt/camera_view.hpp

@@ -19,7 +19,7 @@ class CameraView
     // float calc_resolution_scale(const npz &camview) const;
     unsigned int create_framebuffer();
 
-    glm::mat4 wc2img, projection, current_frame_view_matrix, next_frame_view_matrix;
+    glm::mat4 cc2img, projection, current_frame_view_matrix, next_frame_view_matrix;
     glm::vec3 position;
     double fx, fy, cx, cy;
     float buffer_over_image; // should be >= 1

infinigen/datagen/customgt/glsl/spine.geom

@@ -3,10 +3,7 @@
 layout (lines_adjacency) in;
 layout (line_strip, max_vertices = 2) out;
 
-uniform mat4 wc2img;
-
-out vec3 normal;
-out vec3 interp_pos_wc;
+uniform mat4 cc2img;
 
 in VS_OUT {
     vec3 pos_wc;
@@ -19,7 +16,7 @@ in VS_OUT {
 } gs_in[];
 
 vec2 proj(vec3 v){
-    vec4 h = wc2img * vec4(v, 1.0);
+    vec4 h = cc2img * vec4(v, 1.0);
     return vec2(h[0] / abs(h[2]), h[1] / abs(h[2]));
 }
 
@@ -38,10 +35,10 @@ bool is_frontfacing(vec3 v1, vec3 v2, vec3 v3){
 }
 
 void main() {
-    vec3 v1 = gs_in[0].pos_wc; // A
-    vec3 v2 = gs_in[1].pos_wc; // B
-    vec3 v3 = gs_in[2].pos_wc;
-    vec3 v4 = gs_in[3].pos_wc;
+    vec3 v1 = gs_in[0].pos_cc; // A
+    vec3 v2 = gs_in[1].pos_cc; // B
+    vec3 v3 = gs_in[2].pos_cc;
+    vec3 v4 = gs_in[3].pos_cc;
 
     bool draw_boundary = (is_frontfacing(v1, v2, v3) != is_frontfacing(v2, v1, v4));
 

infinigen/datagen/customgt/glsl/wings.frag

@@ -2,14 +2,10 @@
 
 #version 440 core
 
-uniform vec3 cameraPos;
-
 uniform int object_index;
 
-in vec3 interp_pos_wc;
 in vec3 interp_pos_cc;
 in vec3 interp_pos_cc_next;
-in vec3 normal;
 in vec3 cc_normal;
 in float tri_area;
 in float px_area;
@@ -27,36 +23,42 @@ layout (location = 5) out ivec4 tag_segmentation;
 layout (location = 6) out ivec4 instance_segmentation;
 layout (location = 7) out vec4 geo_normal;
 
+mat4 cv_to_sn_convention = mat4(
+    vec4(1., 0., 0., 0.),
+    vec4(0., -1., 0., 0.),
+    vec4(0., 0., -1., 0.),
+    vec4(0., 0., 0., 1.));
+
 /*
 // No longer used
 layout (location = 3) out vec4 faceSize;
 layout (location = 4) out vec4 pixelSize;
 */
 
 void main() {
-	vec3 updated_normal = normal;
-	if (dot((cameraPos - interp_pos_wc), updated_normal) < 0){
-		updated_normal = updated_normal * -1;
-	}
-
-	rasterized_cc = vec4(interp_pos_cc, 1.0);
-	if (has_flow > 0.99)
-		next_rasterized_cc = vec4(interp_pos_cc_next, 1.0);
-	else
-		next_rasterized_cc = vec4(0.0, 0.0, -1.0, 1.0);
-	tag_segmentation = ivec4(tag, 0);
-	instance_segmentation = ivec4(instance_id[0], instance_id[1], instance_id[2], 1);
-	object_segmentation = ivec4(object_index, 0, 0, 1);
-
-	geo_normal = vec4(normalize(-cc_normal), 1.0);
-	rasterized_occ_bounds = ivec4(0, 0, 0, 1);
-
-	rasterized_face_id = vec4(face_id, 1.0);
-
-	/*
-	// No longer used
-	faceSize = vec4(tri_area, 0.0, 0.0, 1.0);
-	pixelSize = vec4(px_area, 0.0, 0.0, 1.0);
-	*/
+
+    rasterized_cc = vec4(interp_pos_cc, 1.0);
+    if (has_flow > 0.99)
+        next_rasterized_cc = vec4(interp_pos_cc_next, 1.0);
+    else
+        next_rasterized_cc = vec4(0.0, 0.0, -1.0, 1.0);
+    tag_segmentation = ivec4(tag, 0);
+    instance_segmentation = ivec4(instance_id[0], instance_id[1], instance_id[2], 1);
+    object_segmentation = ivec4(object_index, 0, 0, 1);
+
+    if (dot(interp_pos_cc, cc_normal) < 0){
+        geo_normal = cv_to_sn_convention * vec4(cc_normal, 1.0);
+    } else {
+        geo_normal = cv_to_sn_convention * vec4(-cc_normal, 1.0);
+    }
+    rasterized_occ_bounds = ivec4(0, 0, 0, 1);
+
+    rasterized_face_id = vec4(face_id, 1.0);
+
+    /*
+    // No longer used
+    faceSize = vec4(tri_area, 0.0, 0.0, 1.0);
+    pixelSize = vec4(px_area, 0.0, 0.0, 1.0);
+    */
 
 }