SECOND: Sparsely Embedded Convolutional Detection

SECOND: Sparsely Embedded Convolutional Detection Yan Yan Yuxing Mao Bo Li

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

SECOND: Sparsely Embedded Convolutional Detection 2018 Sensors 10.3390/s18103337 yan2018-second-sparsely-embedded-convoluti yan2018-second-sparsely-embedded-convoluti.txt, the project's staged copy bc495a0da9775361b2a517c868886856670604ff cc-by computer-science

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them 3D object detection from LiDAR point clouds, and the speed and orientation-accuracy limits of voxel-based convolutional detectors Quotation withheld because this record is not cleared for redistribution. how to keep the information a 3D voxel representation preserves while removing the inference cost and the orientation error that come with it Quotation withheld because this record is not cleared for redistribution. sparse convolution with a GPU rule-generation algorithm, a sine-error angle loss with a separate direction classifier, and ground-truth sampling augmentation Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. an ill-posed regression target: two orientation values that describe the same box but generate maximal loss against each other Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. objects and their enclosed points are collected once into a database, then pasted into other point clouds during training, which multiplies the labelled objects per scene and simulates each object in new surroundings sampling ground truths from a database Quotation withheld because this record is not cleared for redistribution. pasted objects are removed when they intersect existing ones, so the augmentation cannot produce a scene that could not physically exist collision rejection after sampling Quotation withheld because this record is not cleared for redistribution. each ground truth and its points are transformed independently instead of the whole cloud being transformed together, so the augmentation varies objects relative to their scene per-object rather than per-scene noise Quotation withheld because this record is not cleared for redistribution. performance is lower for pedestrians and cyclists and for bird's-eye-view detection, which the paper states in its own conclusion Quotation withheld because this record is not cleared for redistribution. pedestrians and cyclists give fewer points per instance and higher instance density, so they are more easily confused with clutter Quotation withheld because this record is not cleared for redistribution. test-server access was limited, so the larger model was evaluated on the test set only and the validation assessment covers the smaller model alone Quotation withheld because this record is not cleared for redistribution. the focal loss, carried from single-stage 2D image detection into 3D point-cloud detection to handle the same foreground-background imbalance Quotation withheld because this record is not cleared for redistribution. the same detector with dense convolution instead of sparse, to attribute the speed gain eliminates: a speed gain produced by the rest of the detector rather than by sparse convolution · outcome: held Quotation withheld because this record is not cleared for redistribution. the angle encoding used by AVOD and ComplexYOLO, run as the comparison for the new angle loss eliminates: a new angle loss credited with an improvement that the encoding used by the baselines would also have produced · outcome: held Quotation withheld because this record is not cleared for redistribution. a physical-plausibility check on the augmented scenes eliminates: a detector trained on augmented scenes containing objects that could not physically coexist, which would teach it geometry no sensor will ever show it · outcome: held Quotation withheld because this record is not cleared for redistribution. a validation-set assessment of the larger model eliminates: over-fitting of the larger model to the test set - the validation-set assessment that would show it exists only for the smaller model, because access to the test server was limited · outcome: absent-declared-unreported Quotation withheld because this record is not cleared for redistribution. sparse convolution sparse convolution Quotation withheld because this record is not cleared for redistribution. angle regression angle regression Quotation withheld because this record is not cleared for redistribution. corner encoding corner encoding Quotation withheld because this record is not cleared for redistribution. direction classifier direction classifier Quotation withheld because this record is not cleared for redistribution. ground truths ground truths Quotation withheld because this record is not cleared for redistribution. anchors anchors Quotation withheld because this record is not cleared for redistribution. VoxelNet VoxelNet Quotation withheld because this record is not cleared for redistribution. KITTI KITTI Quotation withheld because this record is not cleared for redistribution. 'orientation error' was measuring the ANGLE and not the BOX - two angles describing one box produced maximal loss Quotation withheld because this record is not cleared for redistribution. the new loss treats boxes with opposite directions as the SAME by design, and a separate classifier recovers the direction Quotation withheld because this record is not cleared for redistribution. the direction target is defined by a sign test on yaw: positive above zero, negative otherwise Quotation withheld because this record is not cleared for redistribution. 'state of the art' here is dataset-relative: 90% AP for 2D car detection against 15% for 3D image-based Quotation withheld because this record is not cleared for redistribution. 'real-time' is stated as a run time, not a claim - 20 fps for the large model and 40 for the small Quotation withheld because this record is not cleared for redistribution. extends → VoxelNet, which combines raw-point and voxel feature extraction end to end note: keeps the voxel approach and replaces the dense 3D convolution that made it too slow Quotation withheld because this record is not cleared for redistribution. critiques → prior angle encodings - corner, direct and vector note: each is rejected for a different failure, and the replacement is a loss that changes what is being measured rather than how well Quotation withheld because this record is not cleared for redistribution. builds-on → the focal loss introduced for single-stage detection Quotation withheld because this record is not cleared for redistribution. anchors —contrasted-with→ ground truths anchors ground truths differs in: count - about 70,000 anchors against a few ground truths each yielding 4 to 6 positives · strength of the relation: 0.95 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span contrasts the NUMBERS of anchors and ground truths; contrasted-with is loose. Quotation withheld because this record is not cleared for redistribution. sparse convolution gives a large speed gain over dense convolution at slight cost in performance Quotation withheld because this record is not cleared for redistribution. the new angle loss outperforms the vector encoding used by two prior detectors Quotation withheld because this record is not cleared for redistribution. the method reaches state-of-the-art results using LiDAR alone Quotation withheld because this record is not cleared for redistribution. performance remains lower for pedestrians and cyclists than for cars Quotation withheld because this record is not cleared for redistribution. angle regression angle regression Quotation withheld because this record is not cleared for redistribution. direction classifier direction classifier Quotation withheld because this record is not cleared for redistribution. 90% AP for 2D car detection against 15% for 3D image-based car detection — None Quotation withheld because this record is not cleared for redistribution. a factor-of-4 training speedup and factor-of-3 inference speedup against dense convolution — None Quotation withheld because this record is not cleared for redistribution. about 0.025 s per frame for the small model on a GTX 1080 Ti — None Quotation withheld because this record is not cleared for redistribution. 20 fps for the larger model and 40 fps for the smaller — None Quotation withheld because this record is not cleared for redistribution. about 70,000 anchors per point cloud against a few ground truths giving 4 to 6 positives each — None Quotation withheld because this record is not cleared for redistribution. average precision (AP) average precision (AP) Quotation withheld because this record is not cleared for redistribution. orientation orientation Quotation withheld because this record is not cleared for redistribution. anchors anchors Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.3390/s18103337.