Utils/data_processing/docs/CAMERA_AND_RENDERING.md

Camera Placement, Rendering, and Post-processing

This guide describes the public PhysInOne workflow for placing cameras, creating a Level Sequence, rendering EXR passes, and converting those passes into the dataset layout.

1. Prerequisites

Install Unreal Engine 5.5.4 and enable these project plugins:

• Python Editor Script Plugin
• Editor Scripting Utilities
• Movie Render Queue

The project must contain BP_CameraTrajectory and the Movie Render Queue configuration assets referenced by render/render_config.bat.

Install post-processing dependencies in a normal Python environment:

python -m pip install -r requirements-render.txt

The camera scripts run inside Unreal Engine, not in the normal Python environment. The moving-camera script imports NumPy, so NumPy must also be available to Unreal's Python interpreter.

2. Configure BP_CameraTrajectory

Place BP_CameraTrajectory at the center of the physical event. Its rotation must be zero and its X, Y, and Z scales must be identical. Use the smallest uniform scale that still lets every camera capture the entire event.

Field in the Unreal panel	Meaning
Num	Number of managed static cameras named CineCamera_0, CineCamera_1, and so on. Choose the count required by your release protocol; 7 in the reference panel is only an example.
Time	Sequence duration in seconds. The script calculates the authoritative frame count as round(Time × FPS).
Random Sample Direction	Randomly chooses clockwise or counterclockwise movement.
Sample Clockwise	Used only when Random Sample Direction is disabled.
Random Hemisphere	Randomly chooses the upper or lower hemisphere.
Upper Hemisphere	Used only when Random Hemisphere is disabled.
Longtitude Range	Longitude span in degrees. The property is intentionally spelled LongtitudeRange in the current Blueprint/API.
Longitude Angles / Latitude Angles	Output arrays written by the moving-camera generator. Do not edit them manually.
Points Array	Not consumed by the supplied Python scripts.
Total Frames	Not consumed by the supplied Python scripts; Time × FPS is authoritative.

The default frame rate is 30 FPS. Redirect and laser phenomena that require faster motion sampling use 60 FPS.

3. Generate static cameras

Open the target level in Unreal Editor. In the Unreal Output Log, run:

py "D:/path/to/data_processing/camera/generate_static_cameras.py" --focal 18

Optional arguments:

Argument	Default	Description
--focal	18	Focal length in millimeters.
--aperture	10	Camera aperture.
--radius	100	Base sphere radius in Unreal centimeters before Blueprint scale.
--min-latitude	30	Minimum absolute sampling latitude.
--max-latitude	60	Maximum absolute sampling latitude.
--deterministic	off	Removes the stable level-specific longitude offset.

The script replaces only cameras whose labels match CineCamera_<integer>. It does not delete unrelated CineCamera actors or CineCamera_Moving.

Inspect every generated camera in the viewport. Move or resize BP_CameraTrajectory and rerun the script if an important object is occluded or outside the frame.

4. Generate the moving camera and Level Sequence

After approving the static cameras, run:

py "D:/path/to/data_processing/camera/generate_camera_sequence.py" --focal 18

Useful arguments:

Argument	Default	Description
--focal	18	Focal length in millimeters.
--aperture	10	Camera aperture.
--radius	100	Base trajectory radius in centimeters.
--latitude-range	60	Maximum latitude magnitude.
--path-type	random	One of random, arc, sine, or loop.
--seed	level-derived	Explicit seed for a different reproducible path.
--output-root	project Rendered/Videos	Optional camera-metadata root.

For a canonical level such as:

/Game/PhysInOne/Scenes/Val/DoublePhysics/ExampleScene.ExampleScene

the script creates or replaces:

/Game/PhysInOne/Trajectories/Val/DoublePhysics/ExampleScene_trajectory.ExampleScene_trajectory

and writes metadata to:

<ProjectRoot>/Rendered/Videos/Val/DoublePhysics/ExampleScene_trajectory/

The generator replaces the sequence at that exact path and replaces only the actor labeled CineCamera_Moving. It removes old camera metadata files but never removes RGB, depth, segmentation, or other rendered-data directories.

The camera stage is the sole owner of:

transforms_train.json
transforms_val.json
transforms_test.json

It chooses two approximately opposite static cameras for validation, uses the first half of their timeline for validation, uses the first half of the other static cameras for training, and uses the second half of all static cameras for testing. The level-derived seed makes this choice reproducible. Post-processing preserves these files exactly.

Open the generated sequence and inspect the complete Camera Cuts track. To try a different trajectory, rerun with a different --seed. Reduce Longtitude Range if the camera crosses walls or other occluders. Once a path is approved, do not move BP_CameraTrajectory or the static cameras without regenerating the sequence.

5. Configure headless rendering on Windows

Copy the example configuration:

copy render\render_config.example.bat render\render_config.bat

Edit these local paths in render_config.bat:

set "UE_CMD=C:\Program Files\Epic Games\UE_5.5\Engine\Binaries\Win64\UnrealEditor-Cmd.exe"
set "UPROJECT=D:\PhysicBenchmark\PhysInOne.uproject"
set "DDC_LOCAL=D:\UnrealCache\Local"
set "DDC_SHARED=D:\UnrealCache\Shared"

Also verify all nine Movie Render Queue configuration paths: Train, Val, and Test for SinglePhysics, DoublePhysics, and TriplePhysics. The public project stores these .uasset files in:

PhysicBenchmark/Content/PhysInOne/Scripts/

This corresponds to /Game/PhysInOne/Scripts/ inside Unreal Engine. Do not rename or relocate that directory: render_config.bat refers to these package paths directly. The scripts alone cannot recreate the render-pass settings.

Render a list of levels:

render\render_levels_windows.bat -f examples\level_list.txt

Render one level:

render\render_levels_windows.bat -l "/Game/PhysInOne/Scenes/Val/DoublePhysics/ExampleScene.ExampleScene"

The renderer writes render/render_batch.log and records successful levels in render/complete_render_list.log. It skips paths already in the completion file and returns a nonzero exit code if any level fails. It does not terminate other Unreal Editor processes.

6. Convert one completed render directory

Preview-safe conversion keeps the source EXRs:

python postprocess/exr_to_dataset.py \
  "D:/PhysicBenchmark/Rendered/Videos/Val/DoublePhysics/ExampleScene_trajectory"

Delete each source EXR only after its converted file is written and verified:

python postprocess/exr_to_dataset.py \
  "D:/PhysicBenchmark/Rendered/Videos/Val/DoublePhysics/ExampleScene_trajectory" \
  --delete-source-exr

Select passes or worker count when necessary:

python postprocess/exr_to_dataset.py RENDER_DIR \
  --channels rgb,depth,seg \
  --workers 8 \
  --jpeg-quality 95

By default, conversion also creates points3d.ply. It back-projects the first RGB/depth frame from every static camera listed by blender_CineCamera_<index>.json, merges the views in world coordinates, and uniformly samples at most 100,000 colored points. CineCamera_Moving is not used. The default valid-depth range is (0, 20) meters.

Adjust the point-cloud limits when necessary:

python postprocess/exr_to_dataset.py RENDER_DIR \
  --depth-threshold 20 \
  --max-points 100000 \
  --ply-seed 0

Regenerate only the point cloud from existing JPEG/NPZ outputs:

python postprocess/point_cloud.py RENDER_DIR \
  --depth-threshold 20 \
  --max-points 100000

Use --skip-ply only when intentionally converting a channel subset that does not include both RGB and depth. In render_processor.py, --force-ply replaces an existing valid PLY; otherwise a valid existing file is preserved.

Output format:

ExampleScene_trajectory/
├── CineCamera_0/
│   ├── rgb/0000.jpg
│   ├── depth/0000.npz       # key: depth
│   └── seg/0000.npz         # key: seg
├── CineCamera_Moving/
│   ├── rgb/
│   ├── depth/
│   └── seg/
├── blender_CineCamera_0.json
├── points3d.ply             # all static views, first frame only
├── transforms_train.json
├── transforms_val.json
└── transforms_test.json

7. Monitor multiple scenes

Use canonical UE level paths and let the processor infer rendered directories:

python postprocess/render_processor.py \
  --level-list examples/level_list.txt \
  --project-root "D:/PhysicBenchmark" \
  --watch \
  --interval 60 \
  --delete-source-exr \
  --log-dir render_processor_logs

Alternatively, provide a text file containing one rendered directory per line:

python postprocess/render_processor.py \
  --render-list rendered_directories.txt \
  --watch

Without --watch, the processor performs one scan and exits. A pending render produces exit code 2. Progress is flushed to disk in real time:

render_processor_logs/
├── progress.log
├── success_scenes.tsv
└── failed_attempts.tsv

The success file is resumable: a later invocation with the same log directory skips scenes already verified as complete.

8. Important safety behavior

• Camera scripts only delete specifically managed camera actors.
• Regenerating a sequence replaces only the inferred scene sequence asset.
• Camera metadata cleanup never deletes rendered media directories.
• EXRs are retained unless --delete-source-exr is explicitly supplied.
• Post-processing never samples views or rewrites transforms_*.json.
• A scene is not marked complete until points3d.ply is valid, unless --skip-ply was explicitly supplied.
• No script contains or requests a Hugging Face or GitHub access token.