Towards benchmarking Western Bluebird detection in the wild

Estela Monserrat Arriaga Santana; Julian Rosas Scull; Ibeth P. Alarcón; Bibiana Montoya; Aylin Sosa Mejía; Hugo Jair Escalante

cs.CVOct 6, 2026

Towards benchmarking Western Bluebird detection in the wild

Authors: Estela Monserrat Arriaga Santana, Julian Rosas Scull, Ibeth P. Alarcón, Bibiana Montoya, Aylin Sosa Mejía, Hugo Jair Escalante

Organizations: Universidad Nacional Aut´onoma de M´exico, Mexico · Universidad Aut´onoma de Tlaxcala, Mexico · The University of Texas at El Paso, USA and INAOE, Mexico

Abstract

Bird monitoring in natural environments is challenging due to the small size of some species of birds relative to the scene, background clutter, variability in illumination, and the observers' viewpoint. Progress is further limited by the scarcity of large-scale, realistic datasets, which are essential for understanding behavioral patterns. To address this gap, we introduce a new benchmark dataset for the detection and segmentation of Western bluebirds (Sialia Mexicana), comprising over 6,000 labeled images from 41 recording sessions. The dataset features high-resolution (4K) in-the-wild images in which birds occupy only a small fraction of the image. We evaluated supervised detectors, open-vocabulary models under zero-shot and fine-tuned settings, and segmentation approaches. Supervised detectors remain the most reliable overall, with Faster R-CNN achieving the highest detection mAP and RT-DETR offering the best precision-recall trade-off. Open-vocabulary models perform poorly in zero-shot settings; however, fine-tuning substantially improves their performance, with YOLO-World becoming competitive with supervised methods and achieving the highest precision, F1-score, and mAP@0.5. For segmentation, supervised methods significantly outperform Grounded-SAM and SAM 3: Mask R-CNN achieves the highest mask mAP, while YOLOv8-Seg provides the best precision and fastest inference. A diagnostic analysis further shows that failures are not explained by object size alone, but by a combination of apparent scale, brightness, contrast, clutter, blur, crowding, and recording-session variation. Overall, our findings highlight the difficulty of zero-shot bird detection in cluttered ecological scenes and underscore the importance of domain adaptation in small-object settings.

Figures & tables

Figure 1 : Close-up of a male Western Bluebird, the species considered in this study.

Figure 2 : Sample images from the dataset.

Figure 3 : Example from the proposed bird detection dataset. (a) Original image captured under challenging real-world conditions. (b) Ground-truth bounding box annotations. (c) Instance-level segmentation masks. (d) Combined visualization of annotations.

Figure 4 : Qualitative results on the test set. (a) True positive from YOLOv8. (b) False negative from YOLO-World (FT). (c) False positive from OWL-ViT (FT). (d) Over-detection from OWL-ViT (ZS). (e) Under-detection from YOLO-World (ZS). (f) Successful segmentation from YOLOv8-Seg. Ground truth is shown in green, correct predictions in blue, and incorrect predictions in red.

Paradigm	Model	mAP	mAP@0.5	Precision	Recall	F1	FPS
Sup. / FT	YOLOv8	0.3804	0.8428	0.8977	0.7338	0.8076	13.49
Sup. / FT	YOLO26	0.3633	0.8772	0.9704	0.8852	0.9259	13.78
Sup. / FT	Faster R-CNN	0.4662	0.9026	0.4500	0.8987	0.5997	7.36
Sup. / FT	RT-DETR	0.4170	0.9340	0.9120	0.9210	0.9170	5.50
Op-v. / ZS	OWL-ViT	0.0070	0.0189	0.0376	0.4609	0.0695	3.34
Op-v. / FT	OWL-ViT	0.1264	0.2770	0.2002	0.2546	0.2241	4.18

Table 1 : Detection benchmark results on the test set. mAP denotes box AP.

Paradigm	Model	mAP	mAP@0.5	Precision	Recall	F1	mIoU	Dice	FPS
Sup. / FT	YOLOv8-Seg	0.1820	0.6653	0.7991	0.7136	0.7539	0.6459	0.7823	13.83
Sup. / FT	Mask R-CNN	0.6398	0.9154	0.6701	0.9441	0.7838	0.8488	0.9166	7.75
Prmpt. / ZS	Grounded-SAM	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.91
Op-v. / ZS	SAM 3	0.2229	0.3186	0.2282	0.7252	0.3472	0.1793	0.1931	0.29

Table 2 : Segmentation benchmark results on the test set. mAP denotes mask AP.

Model	COCO-M	COCO-L	$\mathbf{1/32\text{--}1/16}$	$\mathbf{1/16\text{--}1/8}$
YOLO-World FT	0.5692	0.5311	0.5206	0.5329
RT-DETR FT	0.5028	0.4304	0.4404	0.4313
YOLO-World ZS	0.0436	0.1902	0.1148	0.2050
Mask R-CNN FT	0.7223	0.7305	0.7363	0.6894
SAM 3 ZS	0.7373	0.7504	0.7341	0.7159

Table 3: AP-based scale analysis on the error-analysis split. COCO-M and COCO-L report COCO-style AP, while the last two columns report AP over the only populated USB-style relative-scale bins.

Model	Global brightness	Local brightness
YOLO-World FT	Neutral: 85.6%, $\Delta$ =12.1	Dark: 77.7%, $\Delta$ =20.6
RT-DETR FT	Dark: 84.0%, $\Delta$ =15.4	Dark: 77.7%, $\Delta$ =20.6
Mask R-CNN FT	Neutral: 93.7%, $\Delta$ =5.2	Bright: 95.6%, $\Delta$ =2.7
YOLO-World ZS	Bright: 22.8%, $\Delta$ =24.7	Dark: 19.4%, $\Delta$ =26.0
SAM 3 ZS	Neutral: 83.9%, $\Delta$ =11.6	Dark: 86.3%, $\Delta$ =8.0

Table 4 : Brightness difficulty analysis.

Model	Global contrast	Local contrast
YOLO-World FT	Medium: 86.2%, $\Delta$ =8.7	Low: 86.9%, $\Delta$ =7.0
RT-DETR FT	Medium: 84.5%, $\Delta$ =10.9	Low: 85.7%, $\Delta$ =9.3
Mask R-CNN FT	Low: 94.3%, $\Delta$ =4.0	Low: 93.7%, $\Delta$ =5.7
YOLO-World ZS	High: 21.7%, $\Delta$ =20.0	Low: 17.7%, $\Delta$ =38.4
SAM 3 ZS	Low: 87.4%, $\Delta$ =4.2	Low: 81.1%, $\Delta$ =13.9

Table 5 : Contrast difficulty analysis.

Model	Local blur / sharpness
YOLO-World FT	Medium: 87.4%, $\Delta$ =7.1
RT-DETR FT	Medium: 85.1%, $\Delta$ =11.6
Mask R-CNN FT	Blurred: 93.7%, $\Delta$ =6.3
YOLO-World ZS	Blurred: 20.0%, $\Delta$ =30.6
SAM 3 ZS	Blurred: 84.6%, $\Delta$ =10.4

Table 6 : Blur/sharpness difficulty analysis.

Model	Relative size $s_{r}$
YOLO-World FT	Very small: 78.3%, $\Delta$ =21.1
RT-DETR FT	Very small: 83.5%, $\Delta$ =15.2
Mask R-CNN FT	Large: 93.2%, $\Delta$ =6.1
YOLO-World ZS	Very small: 17.4%, $\Delta$ =33.4
SAM 3 ZS	Very small: 85.2%, $\Delta$ =10.8

Table 7 : Object-size difficulty analysis using the relative scale index

s_{r}

Model	Background clutter	Local clutter
YOLO-World FT	Cluttered: 79.4%, $\Delta$ =18.3	Cluttered: 85.0%, $\Delta$ =10.4
RT-DETR FT	Cluttered: 76.7%, $\Delta$ =22.8	Cluttered: 81.7%, $\Delta$ =14.3
Mask R-CNN FT	Moderate: 94.3%, $\Delta$ =4.1	Clean: 96.0%, $\Delta$ =2.3
YOLO-World ZS	Clean: 23.4%, $\Delta$ =24.3	Cluttered: 16.1%, $\Delta$ =31.0
SAM 3 ZS	Cluttered: 86.1%, $\Delta$ =8.7	Clean: 87.4%, $\Delta$ =5.7

Table 8 : Clutter difficulty analysis.

Model	Crowding
YOLO-World FT	2–3 objects: 72.9%, $\Delta$ =19.4
RT-DETR FT	2–3 objects: 87.5%, $\Delta$ =3.1
Mask R-CNN FT	2–3 objects: 89.6%, $\Delta$ =7.9
YOLO-World ZS	2–3 objects: 10.4%, $\Delta$ =23.3
SAM 3 ZS	2–3 objects: 87.5%, $\Delta$ =2.9

Table 9 : Crowding difficulty analysis.

Model	Recall	Rel. size	Bright.	Contrast	Clutter	Blur	Crowding
YOLO-World FT	90.55%	$0.24$	$\mathbf{-1.37}$	$-0.63$	$-1.02$	$0.39$	$-0.25$
RT-DETR FT	90.36%	$-0.04$	$\mathbf{-2.95}$	$-2.80$	$-0.01$	$0.51$	$0.14$
Mask R-CNN FT	96.79%	$-0.00$	$\mathbf{-5.29}$	$-3.93$	$-3.71$	$0.19$	$-0.72$
YOLO-World ZS	31.57%	$0.64$	$-0.38$	$-0.19$	$-0.05$	$-0.62$	$\mathbf{-0.83}$
SAM 3 ZS	90.17%	$0.08$	$\mathbf{-2.63}$	$-2.30$	$-0.03$	$0.25$	$0.03$

Table 10 : Exploratory multivariable analysis with video fixed effects. Values are logistic-regression coefficients for z-scored predictors; bold indicates the most negative coefficient for each model.

Explore similar work

RareSpot+: A Benchmark, Model, and Active Learning Framework for Small and Rare Wildlife in Aerial Imagery
Apr 21, 2026Bowen Zhang, Jesse T. Boulerice, Charvi Mendiratta +4Aerial Imagery Species
University of California, Santa Barbara, 93106, California, United States of America. · 2Smithsonian National Zoo and Conservation Biology Institute, 20008, Washington, DC, United States of America. · 3Stanford University, 94305, California, United States of America. +1
Time-frequency localization of bird calls in dense soundscapes
Jun 9, 2026Simen Hexeberg, Fanghui Tong, Hari Vishnu +1Animal Vocalizations Passive Acoustic Monitoring
Tropical Marine Science Institute, National University of Singapore, Singapore · Dept. of Electrical and Computer Engineering, National University of Singapore, Singapore · School of Computing, National University of Singapore, Singapore
Overhead Wildlife Locator (OWL): Benchmarking Weakly Supervised Learning for Aerial Wildlife Surveys
Jun 11, 2026Isai Daniel Chacón, Zhongqi Miao, Bruno Demuro +9Species Unmanned Aerial Vehicles
Microsoft AI for Good Lab, Redmond, WA, USA · Center for Research and Formation in Artificial Intelligence (Cinfonia), Universidad de los Andes, Bogotá, Colombia · Conservation X Labs, Washington, D.C., USA +3

Apr 21, 2026, cs.CVJ/K move · Enter open · S save

RareSpot+: A Benchmark, Model, and Active Learning Framework for Small and Rare Wildlife in Aerial Imagery

Bowen Zhang, Jesse T. Boulerice, Charvi Mendiratta, Nikhil Kuniyil+3

University of California, Santa Barbara, 93106, California, United States of America. · 2Smithsonian National Zoo and Conservation Biology Institute, 20008, Washington, DC, United States of America. · 3Stanford University, 94305, California, United States of America.+1

Aerial Imagery Species Pixel-Level Annotation Bird Active Learning

Open paper AskPDF arXiv

Automated wildlife monitoring from aerial imagery is vital for conservation but remains limited by two persistent challenges: the difficulty of detecting small, rare species and the high cost of large-scale expert annotation. Prairie dogs exemplify this problem -- they are ecologically important yet appear tiny, sparsely distributed, and visually indistinct from their surroundings, posing a severe challenge for conventional detection models. To overcome these limitations, we present RareSpot+, a detection framework that integrates multi-scale consistency learning, context-aware augmentation, and geospatially guided active learning to address these issues. A novel multi-scale consistency loss aligns intermediate feature maps across detection heads, enhancing localization of small (approx. 30 pixels wide) objects without architectural changes, while context-aware augmentation improves robustness by synthesizing hard, ecologically plausible examples. A geospatial active learning module exploits domain-specific spatial priors linking prairie dogs and burrows, together with test-time augmentation and a meta-uncertainty model, to reduce redundant labeling. On a 2 km^2 aerial dataset, RareSpot+ improves detection over the baseline mAP@50 by +35.2% (absolute +0.13). Cross-dataset tests on HerdNet, AED, and several other wildlife benchmarks demonstrate robust detector-level transferability. The active learning module further boosts prairie dog AP by 14.5% using an annotation budget of just 1.7% of the unlabeled tiles. Beyond detection, RareSpot+ enables spatial ecological analyses such as clustering and co-occurrence, linking vision-based detection with quantitative ecology.

Abstract

Explore similar work

RareSpot+: A Benchmark, Model, and Active Learning Framework for Small and Rare Wildlife in Aerial Imagery

Time-frequency localization of bird calls in dense soundscapes

Overhead Wildlife Locator (OWL): Benchmarking Weakly Supervised Learning for Aerial Wildlife Surveys