Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife

NeurIPS 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Transductive and semantic-dictionary zero-shot · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces WildBox, a drone-video dataset of 237,505 tracked 3D bounding boxes for six African savanna wildlife classes, and benchmarks two open-vocabulary monocular 3D detectors, showing zero-shot 3D collapse and recovery after fine-tuning, with a zebra merged-class curriculum.

Problem

Monocular 3D detection of wildlife from real aerial drone video is unexplored; existing monocular 3D benchmarks are automotive/indoor and aerial wildlife datasets have only 2D labels.

Approach

A model-in-the-loop (WildLIFT proposals plus human review) tracked 3D box dataset in a per-segment scale-normalised camera frame, plus a benchmark showing zero-shot 3D lifting fails while depth dominates residual error, and a merged-zebra curriculum initialisation helps.

Key results

• Zero-shot 3D AP is 0.00 for both OVMono3D-LIFT and DetAny3D across all 2D-input conditions, while open-vocabulary GroundingDINO reaches 50.55 2D AP@50.
• Curriculum-initialised OVMono3D-LIFT fine-tuning reaches 13.17 ± 0.69 macro AP3D and 8.68 ± 0.47 AP-BEV@0.50, versus 9.12 for direct fine-tuning at 15k iterations.
• Depth accounts for 84.7-92.9% of overall NHD after fine-tuning and 99.0-99.6% in zero-shot conditions.

Surprising findings

• Zero-shot 3D AP is 0.00 even when ground-truth 2D box prompts are supplied, isolating the failure to the 3D stage.
• Extra direct fine-tuning iterations (25k) reduce AP-BEV@0.50 below the 15k run, while the curriculum with fewer total iterations improves it.

Limitations

• Evaluation is frame-by-frame and does not use temporal cues or 3D tracking.
• Curriculum benefit is shown only for one merged pair (zebras); the rhino class merges two species and taxonomic granularity is uneven.
• Annotations are per-segment relative-scale, with no absolute metric scale across flights.

Experimental setup

Datasets: WildBox Omni3D KABR

Benchmarks: WildBox validation

Metrics: AP-BEV@0.25 AP-BEV@0.50 AP3D Rel-AP3D 2D AP@50 normalised Hausdorff distance

Baselines: OVMono3D-LIFT DetAny3D GroundingDINO LabelAny3D

Builds on: Cube R-CNN Omni3D DetAny3D GroundingDINO LabelAny3D

Closest papers by meaning

• Reconstructing Animals and the Wild (CVPR 2025 · similarity 0.87)
• BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.86)
• WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 2024 · similarity 0.86)
• Web-Scale Collection of Video Data for 4D Animal Reconstruction (NeurIPS 2025 · similarity 0.85)
• BirdCollect: A Comprehensive Benchmark for Analyzing Dense Bird Flock Attributes (AAAI 2024 · similarity 0.85)
• Generating the Wild: Individual-Consistent Image-to-Video Generation for Wildlife (NeurIPS 2026 · similarity 0.85)