---

name: inference-and-evaluation description: "Run safe preflight and route legacy Frustum PointNets validation, KITTI result writing, and 2D/BEV/3D AP evaluation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Inference and evaluation

Use this route for checkpoint-backed validation, RGB-detection inference, KITTI label output, and offline AP evaluation. Read the linked CLI and result references before launching; the source assumes precomputed frustum data, 2D detector boxes, and a compatible TensorFlow-1 model.

Route

1. Confirm a checkpoint/model/point/channel match with ../training/SKILL.md and validate the source pickle through ../kitti-data-preparation/SKILL.md.
2. Choose --from_rgb_detection only when the input pickle came from detector boxes and --idx_path is available for frame output.
3. Run python scripts/validate_kitti_results.py --help and validate any existing result directory before compiling/running the evaluator.
4. Keep inference output and evaluator results in fresh directories. Do not overwrite a prior benchmark result.

The evaluator reports 2D, bird's-eye-view, and 3D detection AP. Its native binary is a platform-specific build artifact and is not bundled here. Full inference/AP was not executed as verification: CUDA, checkpoints, and KITTI assets are external gates.

Read evaluation workflow, KITTI results, and troubleshooting. For TensorFlow/CUDA or v2 operator failures route to ../runtime-and-custom-ops/SKILL.md.