File size: 2,816 Bytes
cb8518a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | # Evaluation protocol
This document defines the public OVOW-Bench evaluator behavior. Changes that
affect any item below should update the protocol configuration and release
version.
## Inputs
For every selected `sample_id`, the evaluator requires one GT scene and one
scene from every requested method. Supported scene files are `scene.blend`,
`scene.glb`, and `scene.gltf`.
The dataset is supplied separately with `--data_root`. That directory must
contain `gt/<sample_id>/scene.*` and `<model_name>/<sample_id>/scene.*`.
The track manifest is the source of truth for sample selection. A row can set
`geometry_frame_mode` explicitly. Otherwise a numeric `frame` becomes
`exact:<frame>`; rows without either field use `default_frame_mode`.
## Geometry alignment
The release default is `bbox` alignment with `ground_policy=exclude`.
Alternative modes are exposed for ablations but must be reported when used:
- `identity`: no alignment;
- `bbox`: uniform scale and center translation from scene bounding boxes;
- `object_bbox`: scale and translation anchored by the best first-frame object
match;
- `depth_bbox`: depth-median scale with bounding-box translation;
- `depth_camera`: first-camera canonicalization and depth-median scale.
The evaluator stores the applied 4x4 transform for every sample and method.
## Geometry metrics
Scene-IoU compares the union scene boxes under AABB and OBB representations.
Object matching is solved with the Hungarian algorithm using pairwise AABB IoU
as the assignment score. AABB and OBB IoU are then reported for the matched
pairs. Missing and extra objects are penalized with denominator
`max(gt_object_count, pred_object_count)`.
## Rendered metrics
After geometry alignment, GT and prediction scenes are rendered with the GT
camera. Public defaults are 768x432 resolution, 32 render samples, and the
render engine specified in `configs/release_protocol.json`.
- `PL = 100 * mean((RGB_gt / 255 - RGB_pred / 255)^2)`.
- `N-CLIP = 10 * mean(1 - cosine(CLIP(pred), CLIP(gt)))`.
The CLIP encoder is `openai/clip-vit-base-patch32` by default. Features are L2
normalized before cosine similarity. Both metrics are lower-is-better.
## Aggregation and failures
The public evaluator uses an arithmetic mean over the selected sample rows.
All requested methods must cover the same selected rows. Missing geometry,
failed samples, or missing reports cause a non-zero exit by default.
`--allow_partial` exists only for debugging. Partial summaries include explicit
coverage and missing-pair records and are marked `complete: false`.
## Cache validity
A cached sample report is reused only when its fingerprint matches the current
model list, frame mode, alignment, ground policy, Blender/CLIP arguments,
render settings, and input-scene size and modification time.
|