Frustum PointNets

Use this skill for the CVPR 2018 Frustum PointNets release: a TensorFlow-1 pipeline that lifts image 2D detections into 3D frustums, segments frustum points, and estimates amodal 3D boxes. It covers KITTI as the primary workflow and SUN RGB-D as a beta supplement.

Start safely

This is source code, not an installable Python distribution. Work in an isolated legacy environment and treat Python/TensorFlow/CUDA versions as part of the experiment. A reproducible CPU inspection baseline is:

python -m pip install "tensorflow==1.15.5" "numpy==1.18.5" "scipy==1.4.1" "opencv-python-headless==4.5.5.64" "Pillow==8.4.0" "protobuf==3.20.3"
python -c "import tensorflow as tf; print(tf.__version__)"

Use a Python version supported by the selected TensorFlow wheel. For the repository's unmodified source, prefer its documented Python-2.7/TensorFlow- 1.2/1.4 era; the command above is a Python-3.7 inspection baseline, not an exact benchmark environment. Run python sub-skills/runtime-and-custom-ops/scripts/check_legacy_runtime.py --json from the generated skill root before training or inference.

The source was tested with Python 2.7, TensorFlow 1.2/1.4 GPU, and older CUDA. A Python 3.7/TensorFlow 1.15 CPU graph baseline was verified during skill construction, but CUDA and the v2 custom operators were not. Do not treat a visible GPU or CPU graph as proof of full backend support.

Route by task

• Installation, TensorFlow versions, CUDA, compiler or missing .so files: read runtime-and-custom-ops.
• KITTI layout, calibration, detector rows or frustum-pickle generation: read kitti-data-preparation.
• v1/v2 training, hyperparameters, logs, checkpoints or resume: read training.
• Checkpoint inference, KITTI result rows, evaluator build or AP: read inference-and-evaluation.
• SUN RGB-D preparation, one-hot model, result pickle or Python 3D AP: read sunrgbd.

A full path normally composes runtime → data preparation → training → inference and evaluation. Validate every handoff artifact rather than skipping directly to a long native command.

Cross-cutting constraints

• The repository uses direct module imports and unguarded cPickle in several files. A Python-3 run requires a recorded compatibility port.
• v2 imports custom sampling, grouping, and interpolation operators; missing or ABI-incompatible shared objects are a hard block.
• KITTI and SUN RGB-D class maps, box dimensions, pickles, and coordinate conventions are not interchangeable.
• Data downloads, multi-gigabyte conversion, long training, GUI visualization, and benchmark evaluation require explicit resources and are not smoke tests.
• Use bundled validators/preflights from the owning sub-skill. They are non-destructive and do not depend on this repository checkout.

Read cross-cutting troubleshooting for route- level failures. Read provenance before using this skill with another source revision or refreshing it.