ForestPersons: A Large-Scale Dataset for Under-Canopy Missing Person Detection

Deokyun Kim; Jeongjun Lee; Jungwon Choi; Jonggeon Park; Giyoung Lee; Yookyung Kim; Myungseok Ki; Juho Lee

cs.CVMar 3, 2026

ForestPersons: A Large-Scale Dataset for Under-Canopy Missing Person Detection

Authors: Deokyun Kim, Jeongjun Lee, Jungwon Choi, Jonggeon Park, Giyoung Lee, Yookyung Kim, Myungseok Ki, Juho Lee,

Organizations: Autonomous UAV Research Section, ETRI · Kim Jaechul Graduate School of AI, KAIST

Abstract

Detecting missing persons in forest environments remains a challenge, as dense canopy cover often conceals individuals from detection in top-down or oblique aerial imagery typically captured by Unmanned Aerial Vehicles (UAVs). While UAVs are effective for covering large, inaccessible areas, their aerial perspectives often miss critical visual cues beneath the forest canopy. This limitation underscores the need for under-canopy perspectives better suited for detecting missing persons in such environments. To address this gap, we introduce ForestPersons, a novel large-scale dataset specifically designed for under-canopy person detection. ForestPersons contains 96,482 images and 204,078 annotations collected under diverse environmental and temporal conditions. Each annotation includes a bounding box, pose, and visibility label for occlusion-aware analysis. ForestPersons provides ground-level and low-altitude perspectives that closely reflect the visual conditions encountered by Micro Aerial Vehicles (MAVs) during forest Search and Rescue (SAR) missions. Our baseline evaluations reveal that standard object detection models, trained on prior large-scale object detection datasets or SAR-oriented datasets, show limited performance on ForestPersons. This indicates that prior benchmarks are not well aligned with the challenges of missing person detection under the forest canopy. We offer this benchmark to support advanced person detection capabilities in real-world SAR scenarios. The dataset is publicly available at https://huggingface.co/datasets/etri/ForestPersons.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison of two UAV-based person search scenarios. (a) High-altitude views offer wide-area coverage but often fail to detect targets due to canopy occlusion. (b) Low-altitude MAVs provide closer, ground-level views beneath the canopy, improving the chances of spotting missing persons despite vegetation occlusion.

Dataset	Configuration			Dataset Size		Attributes
Dataset	Scenario	Environments	View Point	#Images	#Annotations	Occlusion	Pose
HERIDAL ( Kundid Vasić & Papić, 2022 )	SAR	Forest	Top-down	1,600	3,194	✗	✗
WiSARD ( Broyles et al., 2022 )	SAR	Forest, Maritime	Oblique	44,588	74,204	✗	✗
SARD ( Sambolek & Ivasic-Kos, 2021 )	SAR	Forest	Oblique	1,981	6,532	✗	✓
VTSaR ( Zhang et al., 2025 )	SAR	Urban, Maritime, Forest	Top-down	12,465	19,956	✗	✗
Visdrone ( Zhu et al., 2021 )	Surveillance	Urban	Oblique	10,209	147,747	✓	✗

Table 1: ForestPersons vs. Others. Comparison of ForestPersons with existing UAV-based datasets containing person class annotations.

Figure 2: Overview of ForestPersons composition pipeline. The full process from data collection in forest environments to frame sampling from video sequences, bounding boxes annotation of missing persons, and difficulty-aware dataset splitting.

Figure 3: Visual samples from ForestPersons. Images depicting individuals in diverse poses, occlusion levels, seasons, and forest environments.

Panel of Figure 4 — Figure 4: Annotation statistics of ForestPersons. Instance-level distribution for pose and visibility (Top) and image-level distribution for season, place, and weather (Bottom).

Figure 5: ForestPersons samples by difficulty level. Shown are representative video sequences from the easy, medium, and hard groups. Predicted boxes (green) are shown with confidence scores, and ground-truth boxes (other colors) are labeled as { pose } _ { visibility level }.

Table 2: Adaptation of prior datasets to under-canopy SAR tasks. Performance comparison of Faster R-CNN trained on each dataset and evaluated on two test sets: the dataset’s own test split and the ForestPersons’ test split. (Left) UAV-based SAR datasets; (Right) ground-level person datasets.

	Validation Split				Test Split
Detection Model	AP	$\text{AP}_{50}$	$\text{AP}_{75}$	AR	AP	$\text{AP}_{50}$	$\text{AP}_{75}$	AR
YOLOv3	55.6	91.7	63.2	63.1	50.2	86.5	53.9	58.6
YOLOX	56.8	92.9	65.2	62.5	51.0	89.0	54.4	58.2
YOLOv11	65.3	95.4	76.6	71.5	65.6	93.4	75.6	71.7
RetinaNet	64.1	96.0	75.8	70.4	64.2	93.9	74.4	70.9
Faster R-CNN	64.2	95.6	76.5	69.6	64.4	92.7	75.4	70.0

Table 3: ForestPersons benchmark results. Object detection model performance on validation and test splits of ForestPersons.

Figure 6: Effect of visibility level on detection performance. Detection precision improves as the visibility level increases across pose attributes.

Train Attributes	Standing			All Poses			Summer			Winter			All Seasons
	(a) Pose						(b) Season
Test Attributes	Standing	Sitting	Lying	Standing	Sitting	Lying	Summer	Fall	Winter	Summer	Fall	Winter	Summer	Fall	Winter
YOLOv3	45.3	30.0	32.1	49.3	51.5	47.5	49.7	53.7	25.7	4.5	1.4	54.0	51.1	58.2	50.7
YOLOX	47.3	30.3	31.7	52.2	50.6	47.9	56.8	57.1	17.2	5.5	1.5	60.0	50.0	53.6	56.5
YOLOv11	60.1	44.5	46.0	65.5	65.7	65.1	65.4	65.4	21.6	6.3	1.6	66.9	65.3	72.8	68.0
RetinaNet	57.5	47.2	43.8	62.3	66.3	60.3	63.4	66.3	43.8	14.6	4.7	63.4	66.0	73.2	63.1

Table 4: Impact of various attributes on detection performance in ForestPersons. Each object detection model was trained and evaluated using subsets of train and test data with unique attributes.

Train dataset	AP	AP 50	AP 75	AR
ForestPersons (artifact-free)	61.4	88.4	76.6	69.0
ForestPersons (augmented)	35.8	64.7	33.9	49.5
SARD	23.2	53.5	18.5	34.8
HERIDAL	0.0	0.0	0.0	0.3
WiSARD	40.2	75.2	35.0	50.0

Table 5: Model performance comparison between the Faster-RCNN model trained with motion-blur augmented data, and the model trained with original data with respect to real MAV dataset.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Missing person detection inference in autonomous SAR systems (a) Autonomous Flight: A MAV performs a search mission under the forest canopy. (b) Onboard Inference: Frames captured by the onboard camera are processed locally on the MAV in real time using a lightweight detection model. (c) Edge Server Inference: Frames are transmitted via a commercial 5G network to a remote edge server, where inference is performed using a higher-capacity model.

Methods	Optimizer	Learning rate	Batch size	Weight decay	Epoch
YOLOv3 ( Redmon & Farhadi, 2018 )	SGD	$1\times 10^{-3}$	64	$5\times 10^{-4}$	273
YOLOX ( Ge et al., 2021 )	SGD	$1\times 10^{-2}$	64	$5\times 10^{-4}$	300
YOLOv11 ( Jocher & Qiu, 2024 )	SGD	$1\times 10^{-2}$	16	$5\times 10^{-4}$	100
RetinaNet ( Lin et al., 2017 )	SGD	$5\times 10^{-3}$	16	$1\times 10^{-4}$	12
Faster R-CNN ( Ren et al., 2015 )	SGD	$2\times 10^{-2}$	16	$1\times 10^{-4}$	12
Deformable Faster R-CNN ( Jocher & Qiu, 2024 )	SGD	$2\times 10^{-2}$	16	$1\times 10^{-4}$	12

Appendix

Table 6: Hyperparameter settings for training object detections. Most configurations follow the default setting of MMDetection and detrex.

Figure 8: Precision-Recall curves of baseline object detection models.

Figure 9: Failure cases of object detection models trained on prior UAV-based SAR and Ground-level Person datasets. Green boxes indicate ground-truth bounding boxes, and red boxes represent model predictions. These examples illustrate the limitations of existing datasets in handling under-canopy SAR scenarios.

Figure 10: Success cases of object detection models trained on ForestPersons. Green boxes indicate ground-truth bounding boxes, and red boxes represent model predictions. The models trained with ForestPersons detect the failure case of the models trained with the existing dataset.

Figure 11: Confusion matrix of the object detection model trained with ForestPersons.

Figure 12: Failure cases of object detection models trained on ForestPersons. Green boxes indicate ground-truth bounding boxes, and red boxes represent model predictions. Ground-truth instances with high occlusion or small bounding box size tend to be frequently missed by the detection model.

Figure 13: Detection cases for models trained on standing-only data and all-pose data. Green boxes indicate ground-truth bounding boxes, and red boxes represent model predictions. Each row corresponds to a model, and each column corresponds to the ground truth pose in the test image. Models trained on a specific pose often fail to detect individuals in other poses and sometimes identify incorrect regions as humans.

Figure 14: Detection cases for models trained on summer-only, winter-only, and all-season data. Green boxes indicate ground-truth bounding boxes, and red boxes represent model predictions. Each row shows detection from a model trained on data from a specific season, while each column represents test data from a particular season. Models trained on limited seasonal data show clear seasonal bias when applied to scenes from different seasons, such as failing to detect people or generating inaccurate bounding boxes.

Panel of Figure 15 — Figure 15: Confusion matrices of object detection models trained on summer and winter datasets. (Left) Summer-trained model; (Right) Winter-trained model.

Figure 16: Examples of false positive cases in person-absent situations. Red boxes represent model predictions.

Figure 17: Examples across various visibility levels and poses. Images are grouped by visibility level (rows) and pose (columns), each drawn from distinct scene contexts.

Figure 18: Difficulty Curve across Train, Validation, and Test Splits

	Annotator A	Annotator B	Annotator C	Annotator D	Annotator E	Annotator F
mean IoU	0.8112	0.7936	0.8249	0.8047	0.7749	0.7975
Precision	0.9432	0.9226	0.9298	0.9374	0.8643	0.9366
Recall	0.9181	0.9086	0.9333	0.8552	0.8495	0.8438
F1 Score	0.9305	0.9155	0.9316	0.8944	0.8569	0.8878
True Positive	482	477	490	449	446	443
False Positive	29	40	40	37	70	30

Appendix

Table 7: Bounding Box Inter-Annotator Agreements.

	Annotator A	Annotator B	Annotator C	Annotator D	Annotator E	Annotator F
Pose	0.8963	0.8952	0.8577	0.8976	0.8677	0.9142
Visibility levels	0.6183	0.6310	0.6163	0.6192	0.6188	0.6027

Appendix

Table 8: Attributes Inter-Annotator Agreements.

	Annotator A	Annotator B	Annotator C	Annotator D	Annotator E	Annotator F
Pose	0.8396	0.8346	0.8577	0.8410	0.7952	0.8654
Visibility levels	0.4514	0.4588	0.4574	0.4537	0.4457	0.4271

Appendix

Table 9: Attributes Cohen’s

\kappa

across annotators.

Figure 19: Visual samples from ForestPersonsIR. Compared to the RGB images shown in Fig. 3 , thermal IR imagery simplifies complex visual patterns and reveals clearer cues for detecting missing persons.

	Validation Split				Test Split
Detection Model	AP	$\text{AP}_{50}$	$\text{AP}_{75}$	AR	AP	$\text{AP}_{50}$	$\text{AP}_{75}$	AR
YOLOv3 ( Redmon & Farhadi, 2018 )	71.2	97.4	82.7	77.7	68.9	96.4	80.8	75.3
YOLOX ( Ge et al., 2021 )	73.9	97.7	85.5	77.9	72.5	96.5	84.3	76.4
YOLOv11 ( Jocher & Qiu, 2024 )	84.1	99.0	96.4	86.7	74.0	98.0	86.1	79.0
RetinaNet ( Lin et al., 2017 )	72.7	98.5	85.0	78.3	70.8	97.8	82.2	76.6
Faster R-CNN ( Ren et al., 2015 )	72.1	97.7	84.5	76.8	71.5	97.4	83.7	75.9

Appendix

Table 10: ForestPersonsIR benchmark results. Object detection model performance on validation and test splits of ForestPersonsIR.

Panel of Figure 20 — Figure 20: Prompt template used to enforce structured bounding box outputs from generative MLLMs.

Proprietary MLLMs
Detection Model	Test Split
Detection Model	AP	$\text{AP}_{50}$	$\text{AP}_{75}$	AR
GPT-4o ( Hurst et al., 2024 )	0.0	0.2	0.0	0.6
GPT-5 ( OpenAI, 2025 )	0.1	0.7	0.0	1.4
Gemini 2.5 Flash ( Comanici et al., 2025 )	0.0	0.2	0.0	0.6
Gemini 2.5 Pro ( Comanici et al., 2025 )	2.2	11.5	0.1	8.1

Appendix

Table 11: VLM detection results on ForestPersons. Evaluation of vision–language models (VLMs), including closed-weight, open-weight, and open-vocabulary variants, on the ForestPersons test split.

Ratio of Train Set	100%	90%	80%	70%	60%	50%	40%	30%	20%	10%
AP (Test Split)	65.3	65.3	64.4	64.9	64.2	64.2	63.6	62.5	61.6	60.8

Appendix

Table 12: Performance of Faster R-CNN trained on subsets of frames sampled at different ratios from each clip.

Ratio of Train Set	100%	90%	80%	70%	60%	50%	40%	30%	20%	10%
AP (Test Split)	65.3	65.1	63.7	63.2	62.6	62.0	60.2	59.7	56.2	53.5

Appendix

Table 13: Performance of Faster R-CNN trained on decreasing numbers of training clips.

Table 34

Figure 21: Speed-accuracy trade-off comparison of various object detection models on the ForestPersons. The x-axis represents the inference speed in Frames Per Second (FPS), and the y-axis represents the Average Precision (AP).

Dataset split	mAP	mAP 50	mAP 75	mAR
Validation	62.0	90.8	74.7	69.5
Test	56.7	84.6	66.5	67.3

Appendix

Table 16: The performance of Faster RCNN models for pose classification tasks.

	Validation Split				Test Split
Model (context)	AP	AP 50	AP 75	AR	AP	AP 50	AP 75	AR
Original (no additional context)	64.2	95.6	76.5	69.6	64.4	92.7	75.4	70.0
Weather (FiLM)	64.4	94.6	77.0	69.4	65.2	92.9	77.4	70.2
Place (FiLM)	64.1	94.6	76.2	69.3	65.2	93.0	76.8	70.3

Appendix

Table 17: Performance of Faster R-CNN models with different contextual inputs using FiLM.

Figure 22: Examples of ForestPersons augmented by motion blur and sensor noise . The motion-blur augmentation has the potential to approximate artifacts caused by MAV maneuvers, exposing the model to more realistic motion-induced artifacts.

Figure 23: Predictions of the trained models on the test dataset collected with real MAV. Faster R-CNN trained on existing SAR datasets exhibit inferior detection performance compared to model trained on ForestPersons, primarily due to the domain gap arising from differences in altitude.

Panel of Figure 24 — Figure 24: Synthetic data to depict the extreme situation in the SAR task using generative models. The red box indicates the bounding box conditioned on the generative model. (Left) The generated images created by generative models finetuned by ForestPersons. (Right) The generated images created by non-finetuned generative models.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Multimodal Object Detection Under Sparse Forest-Canopy Occlusion
May 14, 2026Nitik Jain, Mangal KothariInfrared-Visible Image Fusion Object Detection
Robotics & AI, Johns Hopkins University, USA. This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur. · This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur. · Senior Principal Flight Control Engineer, ADASI, EDGE Group, Abu Dhabi, UAE. This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur.
SilvaScenes: Tree Detection and Species Classification from Under-Canopy Images in Natural Forests
Oct 10, 2025David-Alexandre Duclos, William Guimont-Martin, Gabriel Jeanson +7Instance Segmentation Tree Species Classification
Northern Robotics Laboratory, Universit´e Laval, Qu´ebec, QC, G1V 0A6, Canada · D´epartement des sciences du bois et de la forˆet, Universit´e Laval, Qu´ebec, QC, G1V 0A6, Canada · Institut des Sciences de la Forˆet temp´er´ee, Universit´e du Qu´ebec en Outaouais, Ripon, QC, J0V 1V0, Canada
M3GA-Wild: A Large-Scale Dataset and Benchmark for Multi-Modal Multi-session Ground-to-Aerial Place Recognition in Forests
Sep 19, 2026Ethan Griffiths, Maryam Haghighat, Simon Denman +2Visual Place Recognition LiDAR Place Recognition
Queensland University of Technology (QUT) · CSIRO Robotics

May 14, 2026J/K move · Enter open · S save

Multimodal Object Detection Under Sparse Forest-Canopy Occlusion

Nitik Jain, Mangal Kothari

Robotics & AI, Johns Hopkins University, USA. This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur. · This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur. · Senior Principal Flight Control Engineer, ADASI, EDGE Group, Abu Dhabi, UAE. This research was carried out entirely while the author was with the Department of Aerospace Engineering, IIT Kanpur.

Infrared-Visible Image Fusion Object Detection UAV Remote Sensing

Open paper AskPDF arXiv

Reliable detection of humans beneath forest canopy remains a difficult remote-sensing challenge due to sparse, structured, and viewpoint-dependent occlusion. This paper presents a multimodal proof-of-concept pipeline that integrates three complementary approaches: (i) experimental evaluation of LiDAR returns through vegetation to assess the feasibility of active sensing, (ii) visible--thermal image fusion using a multi-scale transform and sparse-representation framework to enhance human saliency, and (iii) synthetic-aperture image formation via Airborne Optical Sectioning (AOS) to suppress canopy clutter. A YOLOv5 detector is fine-tuned on the Teledyne FLIR thermal dataset and evaluated on thermal and fused imagery. Results show that the tested terrestrial LiDAR configuration provides limited penetration for object-level detection, while visible--thermal fusion improves target visibility in low-contrast scenes and AOS enhances ground-plane detection in synthetic forest imagery. The fine-tuned YOLOv5 achieves a mean average precision of

\sim

0.83 on the top three FLIR classes. These findings establish an initial baseline for UAV-deployable search-and-rescue and surveillance systems operating in forested environments, and motivate future work on dedicated forest datasets and real-time multimodal integration.

Abstract

Object Detection

Explore similar work

Multimodal Object Detection Under Sparse Forest-Canopy Occlusion

SilvaScenes: Tree Detection and Species Classification from Under-Canopy Images in Natural Forests

M3GA-Wild: A Large-Scale Dataset and Benchmark for Multi-Modal Multi-session Ground-to-Aerial Place Recognition in Forests