CAMP3D: 3D Benchmarking & Motion Planning

• CAMP3D is defined by two research threads: a richly annotated 3D point cloud benchmark for urban-scale scene understanding and a prospective extension of contact-aware motion planning to 3D.
• The Campus3D benchmark uses UAV photogrammetry over 1.58 km² with a hierarchical labeling system to enable precise semantic and instance segmentation in outdoor environments.
• The motion planning aspect leverages complementarity constraints and ALM optimization to model intentional robot-object contacts, paving the way for 3D contact-aware navigation.

Searching arXiv for “CAMP3D” and “Campus3D” to ground the article in current literature. CAMP3D denotes an ambiguous label in the available source material. In one, much more explicit, usage it corresponds to Campus3D, a large-scale outdoor point cloud benchmark generated from UAV photogrammetry and designed for hierarchical scene understanding in urban-scale reconstruction and segmentation (Li et al., 2020). In a second, looser usage, the term is associated with the 2025 CAMP paradigm for Contact-Aware Motion Planning Among Movable Objects, where the source material states that CAMP is highly relevant to “CAMP3D” because its optimization and contact modeling already generalize from 2D to 3D; this suggests a conceptual extension rather than a separately canonized method name (Wang et al., 5 Feb 2025). In current technical usage, the benchmark interpretation is the fully specified one, whereas the motion-planning interpretation is best understood as a prospective 3D contact-aware planning framework.

1. Terminological scope and disambiguation

The available source material suggests that CAMP3D is not a single stable term with one universally fixed referent. The first source is the paper “Campus3D: A Photogrammetry Point Cloud Benchmark for Hierarchical Understanding of Outdoor Scene” (Li et al., 2020), which explicitly introduces Campus3D as a richly annotated 3D outdoor scene benchmark. The second source is “Contact-Aware Motion Planning Among Movable Objects” (Wang et al., 5 Feb 2025), which formally introduces CAMP, not a standalone system titled CAMP3D, but states that the formulation is highly relevant to CAMP3D because the robot is modeled as a sphere in 3D, movable objects can be represented as convex hulls in 3D, and the same complementarity-based contact modeling can be extended to 3D contact-rich manipulation and navigation.

This distinction matters because the two lines of work address fundamentally different technical problems. Campus3D concerns 3D point cloud acquisition, annotation, hierarchical semantic segmentation, and instance segmentation. CAMP concerns trajectory optimization under intentional contact for mobile robots among movable objects. A common misconception would be to treat them as a single research artifact. The source material instead supports a bifurcated interpretation: an established dataset benchmark on one side, and a planning formalism whose 3D extension is plausible on the other.

2. Campus3D as an outdoor photogrammetry benchmark

Campus3D is presented as a large-scale outdoor point cloud benchmark built specifically to support hierarchical scene understanding for urban-scale reconstruction and segmentation (Li et al., 2020). The motivation is that existing 3D datasets are not ideal for outdoor scene reconstruction: RGB-D datasets are limited by range and sunlight sensitivity, LiDAR datasets often suffer from occlusion and do not provide a complete holistic view of buildings and campus structures, and standard outdoor annotations often emphasize traffic elements rather than reconstruction-oriented urban structures such as buildings, roofs, walls, roads, paths, facilities, and landscape elements.

To address this gap, the dataset is created from UAV photogrammetry over the National University of Singapore campus, spanning about 1.58 km². The resulting raw point cloud contains about 0.94 billion points, 2,530 instances, 24 semantic classes, and 6 regions: FASS, FOE, PGP, RA, UCC, and YIH. These regions differ in architectural style and function and are used for dataset splitting and benchmarking.

The acquisition pipeline follows a standard SfM-MVS photogrammetry pipeline. UAV imagery is collected with GPS coordinates using DJI Phantom 4 Pro cameras with a 1-inch 20 MP CMOS sensor; flight planning uses DJI GS Pro and Pix4D Capture; two routing strategies are used, Grid and Circular; and for tall buildings, multiple circular flights at different heights are used to capture vertical structure. Images and GPS data are then processed with Pix4D using SfM-MVS to generate the 3D point cloud. The source material characterizes this as producing a dense, geographically grounded point cloud with a more complete view than LiDAR-based urban scans.

3. Hierarchical annotation and problem formulation

A central contribution of Campus3D is its hierarchical annotation structure, inspired by the level-of-detail notion in CityGML (Li et al., 2020). Each point receives a path through a class tree rather than a single flat label. The paper’s illustrative example is:

construction -> building -> wall/roof

The hierarchy has five granularity levels $(H=5)$, and each point has one label per level. The semantic system is constructed using two criteria: semantic attribute and geometrical attribute. The distinction is important because classes such as roof and driving_road may both be planar while differing semantically and functionally.

The label set includes several special categories: unclassified, defined as sparse or unrecognized regions retained for completeness; path/stair, for pedestrian-only paths and stairs; driving_road, for vehicle-only roads; artificial_landscape, for man-made landscape such as artificial pools; and others, for rare individual objects that do not form enough instances for a dedicated class. After semantic labeling, the dataset also assigns instance labels to each countable class, so different roof pieces and multiple buildings receive separate instance IDs. The source material emphasizes that this is especially relevant for improving building level-of-detail.

The annotation procedure is designed for consistency. Annotation is done mainly by projecting the point cloud into 2D views; labels are first assigned at coarse levels and then refined to finer levels; 2D polygon partitioning is performed in three orthogonal views and then refined using user-defined rotation angles; all work is done using CloudCompare; and every point is labeled at least three times and verified by different annotators.

The hierarchical learning problem is formalized as learning a map from points to a tuple of labels across levels:

$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$

with goal

$f:\mathcal{X}\mapsto \mathcal{Y}.$

The class hierarchy is modeled as a partially ordered set $(\mathbb{C}, \le_\eta)$, where $c_2 \le_\eta c_1$ means $c_1$ is a superclass of $c_2$, or the same class. A fully consistent solution satisfies

$y_H \le_\eta y_{H-1} \le_\eta \cdots \le_\eta y_1.$

This formulation turns hierarchical segmentation into a constrained multi-label semantic labeling problem.

4. Learning framework, consistency metrics, and sampling

Because conventional accuracy and IoU do not measure hierarchical validity, Campus3D introduces explicit consistency metrics (Li et al., 2020). For a prediction $Y^i=(y_1^i,\ldots,y_H^i)$, the consistency proportion is defined as the maximum fraction of levels that can be matched by some fully consistent path:

$\mathrm{CP}(Y^i)= \frac{ \max\limits_{(y_1,\ldots,y_H)\in \mathcal{Y}^{\mathrm{FC}} \sum_{h=1}^{H}\mathbb{1}(y_h^i = y_h) }{H}.$

For a set of predictions and threshold $$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$0, the consistency rate is

$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$1

The experiments focus especially on $$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$2, the fraction of fully consistent predictions.

The proposed learning framework is a two-stage method comprising multi-task learning (MT) and hierarchical ensemble (HE). In Stage 1, the MT model uses a shared encoder with parallel decoders or classification heads, one for each hierarchy level. PointNet++ is used as the backbone. The loss is

$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$3

where the prediction term is a weighted sum of cross-entropy losses across levels and the consistency regularizer penalizes cases in which child probabilities exceed parent probabilities along valid hierarchical edges. The ablation MT$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$4 removes this consistency term and uses only the prediction loss.

In Stage 2, hierarchical ensemble acts as a post-processing step that enforces strict hierarchical validity by selecting the best root-to-leaf path:

$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$5

Because the output is always a valid tree path, HE guarantees

$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$6

for the selected path predictions. The framework is compared with MC, an independent classifier per level that ignores hierarchy, and MC + HE, which applies hierarchical ensemble after independent classification.

To make learning feasible on the large point cloud, the paper evaluates two sampling strategies: Random block sampling (RBS) and Random centered KNN (RC-KNN). In the reported experiments, RBS with a 12 m × 12 m block and 2048 points performs better and is adopted as the standard sampling method. For benchmark training, the reduced benchmark version Campus3D-reduced is produced by voxel sampling with 0.15 m resolution, and the split is Training: FASS, YIH, RA, UCC; Validation: PGP; Test: FOE.

5. Benchmark tasks and empirical findings

Campus3D provides benchmarks for hierarchical semantic segmentation and instance segmentation (Li et al., 2020). Hierarchical semantic segmentation is evaluated using OA, IoU / mIoU, and CR$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$7. Baseline segmentation models include PointNet++, PointCNN, and DGCNN. Instance segmentation is benchmarked at level $$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$8, where there are enough countable classes and instances, using ASIS and SGPN as baselines and WCov as the metric. The evaluated countable classes are play_field, vehicle, building, and link, while non-countable classes such as natural, path/stair, not vehicle, and facility are excluded.

The main findings are explicitly hierarchical. Methods that exploit hierarchy outperform the flat multi-classifier baseline: MC+HE improves over MC, MT improves over MT$$\mathcal{D} = \{(X^i, Y^i)\}_{i=1}^N,\qquad Y^i \in \mathcal{Y}= C^1 \times C^2 \times \cdots \times C^H$$9, and MT+HE is generally best or tied-best. On the test set, overall accuracy rises from MC: 85.3% at $f:\mathcal{X}\mapsto \mathcal{Y}.$0, 74.0% at $f:\mathcal{X}\mapsto \mathcal{Y}.$1 to MT: 90.7% at $f:\mathcal{X}\mapsto \mathcal{Y}.$2, 75.2% at $f:\mathcal{X}\mapsto \mathcal{Y}.$3. The consistency metric also improves substantially: MT has about 15% higher $f:\mathcal{X}\mapsto \mathcal{Y}.$4 than MC, and MT$f:\mathcal{X}\mapsto \mathcal{Y}.$5 drops by roughly 10% in $f:\mathcal{X}\mapsto \mathcal{Y}.$6 compared with MT. These results are used to argue that consistency-aware hierarchical supervision is effective both semantically and structurally.

The paper also emphasizes geometric ambiguity. Geometrically similar regions such as roof surfaces and roads may be confused by a flat classifier because they share similar geometry but differ semantically. Hierarchical labels provide coarse-to-fine context that reduces this ambiguity; the source material gives the example that a roof is more likely to be correctly placed under construction -> building -> roof under hierarchical supervision.

Sampling choice also matters empirically: RBS consistently outperforms RC-KNN, indicating that structured spatial sampling yields better local context for large-scale point clouds. Among the tested backbones, DGCNN performs best overall, with test mIoU decreasing as the hierarchy becomes more detailed, which the paper characterizes as expected because finer labels are harder and more sparse. For instance segmentation at $f:\mathcal{X}\mapsto \mathcal{Y}.$7, ASIS outperforms SGPN on all reported classes and on average WCov, providing a baseline for future work.

A plausible implication is that Campus3D’s main significance lies not only in scale but in its simultaneous support for multiple tasks at different levels of detail, including conventional semantic segmentation, hierarchy-aware segmentation, and instance segmentation, all from the same UAV-photogrammetric scene representation.

6. CAMP, complementarity constraints, and the prospective “CAMP3D” planner

The second source, “Contact-Aware Motion Planning Among Movable Objects”, introduces CAMP rather than a separate named system called CAMP3D (Wang et al., 5 Feb 2025). The paper addresses environments with movable objects such as chairs, boxes, doors, or other pushable items, and contrasts its approach with classical contact-avoidance-only planners that treat all objects as obstacles and never intentionally touch them. CAMP instead seeks trajectories in which the robot may intentionally contact objects, push them predictably, and use them to increase reachable space or accomplish tasks, especially in NAMO and RAMO.

The framework is a front-end / back-end planner. The front end uses search to generate a discrete sequence of states and segment times. The back end solves a continuous trajectory optimization problem in which the decision vector is

$f:\mathcal{X}\mapsto \mathcal{Y}.$8

with $f:\mathcal{X}\mapsto \mathcal{Y}.$9 the waypoint or state sequence and $(\mathbb{C}, \le_\eta)$0 the segment times. Each segment is modeled as a polynomial, and with fixed endpoints and time allocation the trajectory can be represented as minimum-energy polynomials, reducing optimization dimension.

CAMP’s key technical idea is to model robot-object contact through complementarity constraints. These encode non-penetration, contact force only when in contact, and the condition that force without contact or penetration without force is disallowed. The full problem is explicitly described as an optimization problem with complementarity constraints (OPCC). To solve it, the paper uses the Augmented Lagrangian Method (ALM), with unconstrained subproblems solved by L-BFGS and Lewis-Overton line search. The source material emphasizes that complementarity constraints often violate LICQ and that ALM is robust for such OPCCs.

In simulation, CAMP is compared against GCOPTER. For NAMO, the reported baseline success rates are around 45–60%, whereas CAMP reaches 75–100% depending on scenario, with overall average success about 95% for CAMP and 54.3% for GCOPTER. Success exceeds 90% when there are more than two movable objects. For RAMO, CAMP achieves 100% success for the cylinder case and 95% success for the cube case. Real-world experiments on an omnidirectional mobile robot show that CAMP can generate feasible trajectories for long-duration, long-distance, active-contact tasks, and that changing the objective can alter behavior; the source material gives the example that with $(\mathbb{C}, \le_\eta)$1 the robot may push the object back after moving it, while with $(\mathbb{C}, \le_\eta)$2 it prefers not to do so.

The source material then states why CAMP is relevant to CAMP3D: the formulation already generalizes from 2D to 3D via the configuration dimension $(\mathbb{C}, \le_\eta)$3; the robot is modeled as a sphere in 3D; movable objects can be represented as convex hulls in 3D; and the same complementarity-based contact modeling can be extended to 3D contact-rich manipulation and navigation. This suggests that, in motion-planning usage, CAMP3D is best interpreted as a prospective 3D contact-aware planning variant grounded in CAMP’s complementarity-constrained optimization framework rather than as a separately established, formally introduced method.

7. Position within the research landscape

Taken together, the two source threads place CAMP3D at the intersection of two different 3D research agendas. In the dataset sense, Campus3D addresses a gap between indoor RGB-D datasets and LiDAR-centric outdoor datasets by providing a richly annotated UAV-photogrammetry benchmark for reconstruction-oriented outdoor understanding (Li et al., 2020). In the planning sense, CAMP addresses a gap between collision-free navigation and physically interactive navigation by formulating intentional contact with movable objects as constrained trajectory optimization and by arguing that the same machinery extends naturally to 3D (Wang et al., 5 Feb 2025).

These lines of work are complementary rather than redundant. Campus3D contributes hierarchical scene representation, annotation granularity, and benchmark methodology. CAMP contributes contact-aware trajectory optimization, complementarity-based contact modeling, and ALM-based numerical solution. A plausible implication is that future work using the label “CAMP3D” could combine both sensibilities: richly structured 3D scene understanding on one side and deliberate physical interaction in 3D environments on the other. At present, however, only the Campus3D benchmark is fully specified under a closely matching name, while the planning-related use of CAMP3D remains a forward-looking extension inferred from the stated 3D generality of CAMP.