Source-linked AI summary
Counting Animals in Camera-Traps Image Sequences without Count Labels: Winning Solution to the iWildCam 2021 Challenge
Fagner Cunha, Juan G. Colonna, Eulanda M. dos Santos
TL;DR
The paper addresses sequence-level counting of unique animals when camera-trap bursts create tracking difficulties and count annotations are unavailable. It presents MaxBoxCount, combining species classification with a MegaDetector-based heuristic, and reports first place among 42 challenge teams. The approach is limited to predicting one species per sequence, while deeper tracking-based alternatives were not explored due to time constraints.
Problem
Sequence-level counting of unique individuals is difficult because camera traps provide sparse bursts and training count annotations are unavailable.
Method
MaxBoxCount combines an EfficientNet-B2 species-classification pipeline with a MegaDetectorV4 heuristic that counts high-confidence detections in the image with the sequence maximum.
Results
The approach ranked first in the iWildCam 2021 challenge among 42 entered teams.
Takeaways & Limitations
The solution shows that existing detection and classification components can form a strong baseline for weakly supervised counting in camera-trap sequences.
Takeaways & Limitations
The classification strategy predicts only one species per sequence, and the authors could not explore tracking-based alternatives in depth because of time constraints.
Abstract
from arXiv · showhide
Camera traps have become an essential tool for wildlife monitoring, motivating the development of computer vision methods for the automated extraction of information from these data. While most prior work has focused on species identification, many ecological applications also require estimating the number of unique individuals appearing across short image sequences. This task is particularly challenging because camera traps typically acquire bursts of images at approximately one frame per second, creating large temporal discontinuities that may make conventional multi-object tracking methods unreliable, and because manually collecting individual count annotations is prohibitively expensive. In this work, we describe the winning solution to the iWildCam 2021 Challenge, which introduced a benchmark for counting animals at the sequence level under realistic annotation constraints where count annotations are unavailable for training. Our approach, MaxBoxCount, combines a strong species classification pipeline with a simple yet effective counting heuristic based on MegaDetector detections to estimate the number of unique individuals without requiring count annotations. Code is available at https://github.com/alcunha/iwildcam2021ufam.
1. Introduction
The iWildCam 2021 Challenge addressed sequence-level counting of unique animals when sparse camera-trap bursts and absent count annotations make conventional tracking and supervised counting difficult. The winning MaxBoxCount solution combined species classification with a MegaDetector-based counting heuristic.
- Ecological applications need counts of unique individuals across image sequences in addition to species identification.
- The challenge required sequence-level unique-individual counts rather than independent detection counts, without count annotations for training.Camera traps typically capture bursts at approximately one frame per second, creating temporal discontinuities that may reduce tracking reliability.
- Detection-only counting can overestimate repeated individuals and underestimate individuals appearing in different frames.The task therefore requires correspondence reasoning across the image burst despite sparse temporal information.
- MaxBoxCount combined a strong species-classification pipeline with a MegaDetector-based sequence-level counting heuristic and required no count annotations.The approach was designed for the competition setting and used existing computer vision components for weakly supervised counting.
- The heuristic estimates a sequence count from the number of bounding boxes in the image with the highest number of detections, using confidence threshold τ = 0.8.
2. Materials and Methods
The method combines an EfficientNet-B2 classifier ensemble with MegaDetectorV4 detections to estimate sequence-level animal counts from camera-trap imagery. It uses image and bounding-box classifiers, prediction fusion, and the maximum number of high-confidence detections across a sequence.
- Dataset: The dataset includes WCS camera-trap images, iNaturalist images, and Landsat 8 multispectral imagery for camera locations.The WCS component contains 263,528 images of 206 species from 414 locations across 12 countries.
- MaxBoxCount heuristic: MaxBoxCount uses an EfficientNet-B2 image-burst classifier ensemble and a MegaDetectorV4 bounding-box counting heuristic.The count is based on the image containing the maximum number of detections, which can provide a lower bound under perfect detection.
- Species classification: Image predictions fuse full-image, mirrored full-image, highest-confidence bounding-box, and mirrored bounding-box predictions with weights 0.15, 0.15, 0.35, and 0.35.Sequence species predictions are obtained by averaging predictions from non-empty images.
- Counting heuristic: The count vector uses the maximum number of bounding boxes with confidence above threshold τ across any image in the sequence.When classifiers identify only empty images, the prediction vector contains zeros.
- Limitation: The classification strategy predicts only one species per sequence, although it was a strong baseline and won the iWildCam 2021 challenge.
- Species classifier: The classifiers comprise one full-image EfficientNet-B2 model and one bounding-box model trained on animal crops from MegaDetectorV4 detections.The full-image model was reported as more effective for herds, while the bounding-box model benefits from highlighted animals.
- Training: The models addressed class imbalance with Balanced Group Softmax using four groups defined by training-instance counts.
3. Results
The evaluation uses MCRMSE to capture species-identification and counting errors, and the proposed approach ranked first among 42 teams. Ablations support horizontal flipping, sequence-level prediction averaging, and a 0.8 detection threshold.
- MCRMSE measures both species-identification and counting errors, including false predictions on empty sequences.
- The approach ranked first in the challenge among 42 teams using a private score calculated on 50% of the test set.
- MCRMSE can remain small even when count errors are large because camera-trap sequences usually contain few individuals and incur double penalties.
- The all-zeros baseline performed surprisingly well, underscoring the difficulty of counting under the competition conditions.
- The final heuristic used horizontal-flip augmentation, averaged sequence predictions instead of majority voting, and counted detections above a 0.8 confidence threshold.
4. Conclusion
The paper concludes that a MegaDetectorV4-based heuristic combined with classifier ensembling is a strong baseline for count-label-free sequence counting, despite predicting only one species per sequence. The authors identify tracking animals across images, classifying tracks, and counting them as a more natural but insufficiently explored alternative.
- The proposed heuristic combines MegaDetectorV4 bounding boxes with an ensemble of classifiers for counting without count labels.
- The strategy won the iWildCam 2021 challenge but is limited to predicting one species per sequence.
- The authors consider multi-object tracking, track classification, and track counting a more natural and interpretable alternative.
- DeepSORT and ReID-based variations were not explored in depth because of time constraints.