qhjqhj00 3d-obj-det-seg-eval
Lifting Unlabeled Internet-level Data for 3D Scene Understanding — Chen et al. (2026) (arXiv:2604.01907, 2026)
What this evaluates
Evaluates a model's ability to detect and segment 3D objects in indoor scenes using point cloud inputs. It probes spatial reasoning and instance-level understanding by measuring how well the model generalizes from synthetic internet-scale data to real-world scanned environments.
Datasets
- ScanNet — total ?; splits: test (-1); repo https://github.com/ScanNet/ScanNet
- SceneVerse++ — total ?; splits: train (-1)
Metrics
AP25— range: [0, 1]- Average Precision at Intersection over Union (IoU) threshold of 0.25.
AP50— range: [0, 1]- Average Precision at IoU threshold of 0.50.
AP(primary) — range: [0, 1]- Average Precision averaged across IoU thresholds from 0.25 to 0.75 in 0.05 increments.
Input / output format
Input: 3D point clouds (often cropped to a 3-meter radius around a target object) with optional graph-based segmentation masks.
Output: Bounding boxes and instance segmentation masks for 15 semantic categories.
Scoring recipe
for each scene:
preds = model.predict(point_cloud)
for each pred_box:
match with ground_truth_box
if IoU(pred, gt) >= threshold:
mark as true positive
compute precision-recall curve
interpolate AP at 101 recall levels
return mean AP across categories
Common pitfalls
- Model performance is highly sensitive to the hyperparameters of the graph-based segmentation method (kThresh, segMinVerts) used to generate input segments.
- Distribution shift between synthetic training data (SceneVerse++) and real scanned data (ScanNet) causes significant performance degradation if not carefully managed.
Evidence (verbatim from paper)
The AP is reported as the average over the whole ScanNet.
Citation
@misc{chen2026lifting,
title={Lifting Unlabeled Internet-level Data for 3D Scene Understanding},
author={Chen et al. (2026)},
year={2026},
note={arXiv:2604.01907}
}
- arXiv: 2604.01907