WildCross Benchmark Overview

• WildCross is a cross-modal benchmark offering RGB images, semi-dense depth, lidar, and surface normal data for natural environment perception.
• It addresses the gap in robotics datasets by focusing on forests and trails with challenges like reverse revisits and strong viewpoint changes.
• The benchmark supports visual, lidar, and cross-modal place recognition along with metric depth estimation using rigorous evaluation metrics.

WildCross is a large-scale cross-modal benchmark for place recognition and metric depth estimation in unstructured natural environments. It was introduced to address the mismatch between the predominantly urban or indoor focus of earlier robotics datasets and the conditions faced by field robots in forests, trails, and vegetation-heavy terrain. WildCross comprises over 476K sequential RGB frames aligned with semi-dense depth and surface normal annotations, accurate 6DoF poses, and synchronized dense lidar submaps, and it supports visual place recognition (VPR), lidar place recognition (LPR), cross-modal place recognition (CMPR), and monocular metric depth estimation (Knights et al., 2 Mar 2026). Subsequent work has used it both for expanded cross-modal analysis in natural environments (Hall et al., 10 Jun 2026) and as an evaluation ground for depth-aware VPR methods in forests (Nedov et al., 11 Jun 2026).

1. Motivation, scope, and relation to earlier datasets

WildCross was created to fill a gap in robotics benchmarking: major public datasets such as KITTI and Oxford RobotCar are mainly collected in structured urban scenes, whereas natural environments contain irregular terrain, dense vegetation, narrow trails, strong occlusions, repeated-looking natural structures, large viewpoint changes, reverse revisits, and long-term temporal variation. The benchmark is therefore designed for perception in forests and bushland rather than in lane-structured or building-dominated settings (Knights et al., 2 Mar 2026).

The benchmark is positioned as an extension of Wild-Places. Two additions are central. First, it provides cross-modal alignment between RGB images and lidar submaps with accurate 6DoF pose. Second, it adds semi-dense metric depth and surface-normal annotations for every RGB frame. This changes the scope from lidar-only place recognition to a broader benchmark for VPR, CMPR, and depth estimation in natural environments (Hall et al., 10 Jun 2026).

Relative to prior natural-environment datasets, the benchmark paper distinguishes WildCross from Nordland, Wild-Places, Oxford Forest, RELLIS-3D, and BotanicGarden. Nordland is challenging for VPR but has only one fixed route and no lidar; Wild-Places originally lacks camera data; Oxford Forest lacks camera support for broader cross-modal tasks; RELLIS-3D has limited scale and sequence diversity; and BotanicGarden has synchronized image and lidar data but limited scale, scene diversity, and viewpoint or temporal variation. According to the authors, WildCross is the first large-scale benchmark with sequential depth ground truth in natural environments (Knights et al., 2 Mar 2026).

A common misconception is to treat WildCross as only a forest VPR benchmark. The primary benchmark paper and the expanded analysis both define it more broadly: it is a cross-modal benchmark spanning retrieval and depth estimation, with explicit emphasis on bridging 2D and 3D perception in natural environments (Knights et al., 2 Mar 2026).

2. Dataset composition and annotation pipeline

WildCross spans two Australian natural environments, Venman and Karawatha, with four traversals per environment, for a total of eight traversals collected across 14 months and about 33 km. The sequence naming convention is V-XX for Venman and K-XX for Karawatha. The traversal pattern is deliberately structured: Sequence 01 is the base or forward traversal, Sequence 02 is the reverse trajectory of Sequence 01, Sequence 03 is an alternate extended route, and Sequence 04 is a repeat of Sequence 01. This design introduces forward revisits, reverse revisits, alternate-route overlap, and long-term temporal change (Knights et al., 2 Mar 2026).

The benchmark contains over 476K camera/depth images and 63.3K lidar submaps, with 95.5K revisits in the image stream and 8.7K revisits in submaps. RGB data are obtained from a forward-facing camera, extracted at 15 Hz, and rectified using calibration distortion parameters. Depth annotations are semi-dense rather than fully dense: depth is assigned only at pixels corresponding to visible 3D points, while occluded points, points outside the camera frustum, and points behind the camera are removed (Knights et al., 2 Mar 2026).

Surface normals are derived from local point neighborhoods using points within 0.5 m, with eigendecomposition of the neighborhood covariance; the eigenvector corresponding to the smallest eigenvalue is used as the normal, which is then oriented toward the observer and transformed into the camera frame. Lidar submaps are built from a global accumulated point cloud by taking all points within a 30 m radius and a 1 s time window around the sensor position every 0.5 s along the trajectory. For multiple sessions in the same environment, the global point clouds are aligned using ICP so that all sessions share a common frame (Knights et al., 2 Mar 2026).

The visibility filtering used to generate usable depth labels relies on the generalized hidden point removal (GHPR) operator. The spherical reflection is defined as

$$F(p; \gamma) = \begin{cases} p \| p \|_2^{-1} f(\|p\|_2; \gamma), & \|p\|_2 \neq 0 \ 0, & \|p\|_2 = 0 \end{cases}$$

with

$f(d; \gamma) = d^\gamma,\quad \gamma < 0.$

The benchmark paper explains the intuition as follows: points closer to the camera are mapped farther away, and visible points lie on the convex hull of the transformed cloud. This is the key visibility step used to suppress occluded points in the projected depth maps (Knights et al., 2 Mar 2026).

3. Benchmark tasks, splits, and evaluation criteria

WildCross supports four tasks: VPR, LPR, CMPR, and metric depth estimation. For place recognition, the benchmark introduces a four-fold cross-split design in which sequences with the same index in both environments are held out together for testing and the remaining sequences are used for training. Across the four splits, the expanded analysis reports approximately 298K to 389K training RGB frames, 86.9K to 177.9K test RGB frames, 39.7K to 51.8K training lidar submaps, and 11.5K to 23.6K test lidar submaps (Hall et al., 10 Jun 2026).

Task	Representative methods	Metrics / correctness criterion
VPR	NetVLAD, MixVPR, DINOv2-SALAD, BoQ	Recall@1, Recall@5; correct if within 25 m
LPR	MinkLoc3Dv2, LoGG3D-Net, HOTFormerLoc	Recall@1, Recall@5; correct if within 3 m
CMPR	LIP-Loc with ResNet50, DINOv2 ViT-S, DINOv3 ViT-S	Recall@1, Recall@5; correct if within 25 m
Metric depth estimation	DepthAnythingV2 ViT-S, ViT-B, ViT-L	$\delta_1$, AbsRel, RMSE

For VPR, WildCross does not define positive pairs by distance alone. Positive training pairs are camera poses within 5 m and 15° bearing, while negative pairs are images more than 50 m apart. The benchmark reports both zero-shot and fine-tuned settings, and evaluates both intra-sequence retrieval and inter-sequence retrieval. The 25 m correctness threshold reflects the relocalization granularity used for image-based retrieval in these natural scenes (Knights et al., 2 Mar 2026).

For CMPR, the task is image-to-lidar retrieval. The benchmark paper reports a modified LIP-Loc objective to reduce collapse caused by false negatives, using non-negatives defined as samples within 50 m in the real world. For LPR, positive submap pairs are within 3 m, negative pairs are beyond 20 m, and retrieval is correct if the matched submap is within 3 m, which is stricter than VPR and CMPR because lidar localization is expected to be more precise (Knights et al., 2 Mar 2026).

For metric depth estimation, the benchmark paper uses a task-specific split in which V-01 and K-01 are held out for testing, K-02 is used for validation, and the remaining sequences are used for training. The reported metrics are threshold accuracy $\delta_1$, Absolute Relative Error, and Root Mean Square Error. The expanded analysis gives the formalizations

$$\delta_1 = \frac{1}{N}\sum_{i=1}^{N} \mathbf{1}\left(\max\left(\frac{d_i}{\hat d_i}, \frac{\hat d_i}{d_i}\right) < 1.25\right),$$

$$\mathrm{AbsRel} = \frac{1}{N}\sum_{i=1}^{N}\frac{|d_i-\hat d_i|}{d_i},$$

and

$$\mathrm{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(d_i-\hat d_i)^2},$$

computed only on pixels with known ground truth (Hall et al., 10 Jun 2026).

4. Empirical difficulty and baseline behavior

The main finding across the benchmark papers is that WildCross is substantially harder than common urban benchmarks. In VPR, the expanded analysis reports the best intra-sequence average as fine-tuned BoQ with 63.17% R@1 and 67.91% R@5, and the best inter-sequence average as fine-tuned BoQ with 61.87% R@1 and 66.19% R@5. The same analysis emphasizes that fine-tuning improves all methods substantially, but performance remains far below what comparable methods achieve on datasets such as Pittsburgh and MSLS, where BoQ exceeds 90% R@1 (Hall et al., 10 Jun 2026).

The benchmark paper describes reverse revisits as a central difficulty driver. Sequences V-03 and K-03 have the most reverse revisits and also the worst VPR performance, while reverse sequences such as V-02 and K-02 significantly reduce Recall@1. Because the camera is forward-facing, a revisit from the opposite direction may have very little visual overlap even when the physical location is the same. This makes WildCross a stringent test of viewpoint robustness rather than only a test of appearance invariance (Knights et al., 2 Mar 2026).

LPR performs better than VPR, but it is not saturated. The benchmark paper reports strong intra-sequence LPR results, often above 90% R@1, while inter-sequence average R@1 stays below about 86%. A later WildCross-focused summary explicitly notes that LiDAR place recognition outperformed visual place recognition by about 20–40%, implying that visual features alone are not sufficiently robust for forest place recognition (Nedov et al., 11 Jun 2026). This numerical gap is one of the benchmark’s clearest empirical statements about modality dependence in natural environments.

CMPR is harder still. The benchmark paper reports a best CMPR result of about 51.42% average R@1 for DINOv3-based LIP-Loc, while the expanded analysis reports 52.35% average R@1 and 60.12% average R@5 for LIP-Loc with a DINOv3 backbone. Both presentations support the same conclusion: cross-modal retrieval in forests remains difficult, and stronger backbones alone do not resolve the 2D-to-3D alignment problem (Knights et al., 2 Mar 2026).

Depth estimation exhibits a distinct out-of-domain failure mode. For DepthAnythingV2, zero-shot performance on WildCross is poor, with the expanded analysis reporting for DA2-ViT-S $\delta_1 = 0.284$, AbsRel $= 0.558$, and RMSE $= 7.651$, improving after fine-tuning on WildCross ground truth to $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$0, AbsRel $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$1, and RMSE $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$2. Larger fine-tuned backbones improve further, reaching for DA2-ViT-L $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$3, AbsRel $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$4, and RMSE $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$5 (Hall et al., 10 Jun 2026). The qualitative tradeoff noted in both benchmark papers is that fine-tuning improves global scale and overall correctness but can reduce fine detail, especially for small leaves and fine vegetation structure.

5. WildCross as an evaluation ground for geometry-aware methods

WildCross has already been used to test methods that inject geometric cues into visual place recognition. The paper "Visual Place Recognition in Forests with Depth-Aware Distillation" evaluates a lightweight depth-aware distillation framework, denoted DAD, on WildCross in the zero-shot setting. The method uses a pretrained SALAD checkpoint built on DINOv2, freezes the RGB pathway, and trains only the depth patch embedder and a learnable depth scaling parameter $f(d; \gamma) = d^\gamma,\quad \gamma < 0.$6, with depth input taken from the WildCross depth release generated with Depth Anything V2 (Nedov et al., 11 Jun 2026).

The method defines a frozen teacher descriptor

$f(d; \gamma) = d^\gamma,\quad \gamma < 0.$7

and a student descriptor

$f(d; \gamma) = d^\gamma,\quad \gamma < 0.$8

with a triplet retrieval loss, a cosine-distance alignment loss, and a total loss

$f(d; \gamma) = d^\gamma,\quad \gamma < 0.$9

where $\delta_1$0 is initialized to 0.01 and $\delta_1$1 (Nedov et al., 11 Jun 2026). The point of these constraints is to inject depth while maintaining the pretrained descriptor space.

On WildCross, DAD reaches average intra-sequence 60.82 R@1 and 65.94 R@5, and average inter-sequence 62.69 R@1 and 69.68 R@5. This is a large gain over the frozen appearance-only baseline, which records 45.85/52.24 intra-sequence and 49.38/57.32 inter-sequence, and it also exceeds the RGB fine-tuning baseline in the harder inter-sequence setting, where the fine-tuned baseline records 58.14/63.41 (Nedov et al., 11 Jun 2026). The largest improvements occur in inter-sequence retrieval, with gains of +13.31 absolute in average R@1 and +12.36 in average R@5 relative to the frozen baseline, and +4.55 and +6.27, respectively, relative to the fine-tuned baseline.

The qualitative analysis in that paper identifies three recurring scenarios in which depth helps on WildCross: scale ambiguity, foreground distraction, and distinctive path geometry. The paper also reports inter-sequence R@1 heatmaps showing that improvements are widespread rather than isolated, although not uniform across traversal pairs. This suggests that WildCross is particularly informative for testing whether a method can exploit geometry when appearance is ambiguous (Nedov et al., 11 Jun 2026).

6. Limitations, failure modes, and research implications

WildCross is intentionally difficult, but the benchmark papers also delimit its coverage. The expanded analysis notes that the dataset is geographically limited to two natural environments near Brisbane, Australia; the camera is forward-facing only; depth annotations are semi-dense rather than fully dense; and monocular depth methods still struggle with fine vegetation structure and large unstructured depth variation (Hall et al., 10 Jun 2026). These are benchmark limitations rather than incidental implementation details.

The dominant failure mode exposed by the benchmark is not merely illumination change. The papers repeatedly emphasize repetitive vegetation, weak structural cues, strong viewpoint change, reverse traversal, and large appearance variation across lighting, vegetation growth, and weather. In VPR, RGB-only systems can return visually plausible but incorrect matches; in depth estimation, global scale can improve while fine structure degrades; in CMPR, image-to-point-cloud matching remains only moderately successful even with stronger backbones (Nedov et al., 11 Jun 2026).

The benchmark’s main research implication is that natural-environment perception requires stronger geometric reasoning and tighter coupling between 2D appearance and 3D structure. The expanded analysis suggests improving robustness to reverse revisits and limited field of view, developing depth models that handle thin structures and vegetation without losing fine detail, and expanding natural-environment benchmarks to more sites, seasons, and sensing modalities. The depth-aware distillation study adds more specific directions, including LiDAR-derived geometry, transfer to other backbones, and extension to other unstructured benchmarks (Hall et al., 10 Jun 2026).

In that sense, WildCross functions both as a dataset and as a diagnostic instrument. Its technical contribution is the combination of RGB, semi-dense depth, surface normals, lidar submaps, and accurate 6DoF alignment at scale; its scientific contribution is to show, with consistent empirical evidence, that perception methods successful in structured urban scenes remain brittle in forests and bushland, especially under reverse revisits and cross-modal retrieval (Knights et al., 2 Mar 2026).