Research map / Detection and segmentation / Open-set and few-shot detection
BirdCollect: A Comprehensive Benchmark for Analyzing Dense Bird Flock Attributes
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 BirdCollect, a ~46 GB benchmark of manually annotated dense bird flock images (Demoiselle Cranes in Khichan and 34 species in Bharatpur) and benchmarks crowd counting, density estimation, segmentation, and species classification methods.
Problem
Existing bird datasets contain low-density, crowd-sourced annotations and few birds per image, limiting automated monitoring of dense, species-rich bird flocks in uncontrolled outdoor imagery.
Approach
A large, manually point-annotated, high-density bird flock dataset from two field sites with varied cameras and resolutions, plus SAM-based pseudo segmentation masks, and benchmark results across counting, density, segmentation and fine-grained species classification tasks.
Key results
- Diffuse-Denoise-Count achieved the lowest MAE of 26.18 on the Demoiselle Crane (K,S1) counting benchmark, while P2PNet and M-SFANet had MAE above 145.
- OneFormer reached the highest mIoU of 0.94 on the Bharatpur (B,S2) segmentation set; on K,S1 the best mIoU was 0.76 (OneFormer and SAM with point annotations).
- HERBS achieved 82.11% accuracy on the 35-class species classification protocol, the best among the evaluated fine-grained methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diffuse Denoise-Count instead of CAN | Bird counting · (K, S1) | MAE: 29.37 -> 26.18 | Helped |
| Use Diffuse Denoise-Count instead of CSRNet | Bird counting · (K, S1) | MAE: 28.12 -> 26.18 | Helped |
| Use Diffuse Denoise-Count instead of DMCount | Bird counting · (K, S1) | MAE: 27.44 -> 26.18 | Helped |
| Use Diffuse Denoise-Count instead of CAN | Bird counting · (K, S1) | RMSE: 44.65 -> 41.24 | Helped |
| Use OneFormer instead of Mask2Former | Bird segmentation · (K, S1) | mIoU: 0.24 -> 0.76 | Helped |
| Use OneFormer instead of ClipSeg | Bird segmentation · (K, S1) | mIoU: 0.6 -> 0.76 | Helped |
| Use SAM with point annotations instead of Grounded Dino + SAM | Bird segmentation · (K, S1) | mIoU: 0.49 -> 0.76 | Helped |
| Use OneFormer instead of Mask2Former | Bird segmentation · (B, S2) | mIoU: 0.56 -> 0.94 | Helped |
| Use OneFormer instead of ClipSeg | Bird segmentation · (B, S2) | mIoU: 0.77 -> 0.94 | Helped |
| Use OneFormer instead of SAM + Point Annotation | Bird segmentation · (B, S2) | mIoU: 0.88 -> 0.94 | Helped |
| Use HERBS instead of PMG | Bird species classification · (B, S2) | accuracy: 66.67 -> 82.11% | Helped |
| Use HERBS instead of MC Loss | Bird species classification · (B, S2) | accuracy: 72.10 -> 82.11% | Helped |
Surprising findings
- Segmentation performance of the SAM-based pseudo-labelled setup and some transformer methods varied widely, with ViT-Net reaching only 21.48% accuracy on species classification.
Limitations
- Crowd-counting and density baselines struggle with high occlusion and density variation, leaving the task largely unsolved.
- Segmentation masks for most crane images are SAM-generated pseudo-labels rather than manual masks.
- Species classification is limited by long-tailed class imbalance and few samples for rare species.
Experimental setup
Datasets: BirdCollect (Demoiselle Crane subset K,S1 and Bharatpur species subset B,S2)
Benchmarks: BirdCollect
Metrics: MAE RMSE mIoU accuracy
Baselines: CAN CSRNet P2P-Net M-SFANet DM-Count CrowdFormer Diffuse-Denoise-Count Mask2Former SegFormer ClipSeg OneFormer SAM
Builds on: Segment Anything (SAM) DM-Count Diffuse-Denoise-Count Mask2Former OneFormer
Closest papers by meaning
- WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 2024 · similarity 0.88)
- RealBirdID: Benchmarking Bird Species Identification in the Era of MLLMs (CVPR 2026 · similarity 0.88)
- BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.87)
- MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps (CVPR 2025 · similarity 0.86)
- WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife (NeurIPS 2026 · similarity 0.85)
- Fish-Vista: A Multi-Purpose Dataset for Understanding & Identification of Traits from Images (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.