These Tiny Bees Just Changed the Way We Think About AI Models

Table of Links

Abstract and 1. Introduction

2. Related Works
3. Method
4. Results and Discussion
5. Conclusion and References

4. Results and Discussion

Our initial model selection (Table 1), demonstrates these models’ ability to re-id reed bees over a short 4 hour period.

For comprehensive analysis using all datasets, we chose models with Accuracy > 0.40 (and F1 score > 0.35), seven in total (unshaded in Table 1). Notably, the mnasnet0_75, a very small network with only 1.91M parameters, outperformed many of its larger counterparts. This suggests that smaller models that are practical for their training and inference speed may be useful and underscores the need to consider a range of network architectures for suitability in specific domains. The regnet and mnasnet architectures demonstrated superior performance for our problem.

The comparative assessment of re- and retro-id across our five day experiment for all seven models appears in Table 2 while Figure 2 demonstrates re- and retro-id accuracy of the top performing model (regnet_y_3_2_gf). We found a decline in day-to-day re-id accuracy for all models, consistent with findings from [16]. A similar trend was observed in retro-id for almost all of the models, as demonstrated by the sparkline graphs in the Table 2. Additionally, a two-tailed t-test of the mean accuracies across the days revealed no significant difference in the trends of re- and retro-id average accuracies (t-Stat=0.725, p=0.495) for our reed bee dataset, thus confirming our initial hypothesis.

Retro-id of animal subjects after the fact, and tracking their appearances back through an experiment, can offer a more efficient approach than the conventional chronological re-id. This might be the case when many insects fail to complete a study or when only a few exhibit a target behaviour. For instance, if we were monitoring honeybee hives for bees returning with a specific pollen colour [17] indicative of an invasive weed [1, 10], not many arrivals would be expected to meet our criteria from the thousands of departures each day. Yet if a bee does arrive meeting our criterion, we might like to establish when she departed to estimate her travel time/distance, or even identify from the dance she watched on the honeycomb which direction she is likely to have flown [22, 23]. Trying to id every insect of the large number that leaves the hive, then re-id them again as they return is infeasible and unnecessary. Our retro-id approach would be suitable in this application. Similarly, reed bees share nests and interact semi-socially, making individual identification in natural settings over long periods of time difficult. Using retro-id we only need to annotate data and train algorithms to recognise reed bees that last the full length of an experiment, which would in turn allow us to track complex behaviour and help us understand how often and for how long reed bees forage.

Combining re-id and retro-id provides flexibility to conduct analyses in both temporal directions. Since accuracy decreases gradually as we move away from the training day, if it’s possible to identify an individual of interest midway through an experiment, it would also be advantageous to train models on data from that day. We could then apply both retro- and re-id to potentially reduce the impacts of poor identification performance at temporal extremes.

Domain adaptation [9] is another interesting approach for re-identification that treats changes in appearance over a long period of time as a domain shift. Networks that can find and learn domain-invariant features can be trained as effective re-identification models [16], an approach worth exploring in the future for retro-identification as well.

Overall, our results suggest that retro-identification is a viable and valuable alternative to re-id for longitudinal studies. It is especially helpful when there is an expectation of high subject attrition or late identification of key subjects. Retro-identification is resource-efficient as it allows users to conduct data annotation and model training only on individual experimental subjects of known interest.

5. Conclusion

In this study, we examined the applicability of retro-identification, a conceptual approach and practical, simple technique that identifies past occurrences of experimental subjects in archival image data, in contrast to previous forward-looking re-identification approaches. To demonstrate the approach, we developed a challenging reed bee re-identification dataset of 15 individuals recorded over 5 days. From a pool of 17 transfer learning-based image classification models, 7 were selected for analysis based on their performance on training and test datasets captured over a short 4-hour time frame. Subsequently, we trained the top7 models on Day 1 data, and evaluated their performance on long time frame data in the form of sequences of subsequent data from Day 2 to Day 5 data. And we did the reverse, training the models on Day 5 data and testing our approach by retro-identifying individuals back through the preceding days of data in the archive to Day 1. We found no statistical evidence of differences between re- and retro- identification accuracy, underscoring the potential of retro-identification in longitudinal studies to help manage the sometimes infeasible demands of forward re-identification on large numbers of experimental test subjects, many of whom may ultimately not contribute any useful data to an experimental outcome.

References

[1] Kathryn L. Batchelor, Karen L. Bell, Mariana Campos, and Bruce L. Webber. Can honey bees be used to detect rare plants? Taking an eDNA approach to find the last plants in a weed eradication program. Environmental DNA, 5(6):1516– 1526, 2023. 4

[2] Franziska Boenisch, Benjamin Rosemann, Benjamin Wild, David Dormagen, Fernando Wario, and Tim Landgraf. Tracking All Members of a Honey Bee Colony Over Their Lifetime Using Learned Models of Correspondence. Frontiers in Robotics and AI, 5:35, 2018. 1, 2

[3] Parzival Borlinghaus, Frederic Tausch, and Luca Rettenberger. A Purely Visual Re-ID Approach for Bumblebees (Bombus terrestris). Smart Agricultural Technology, 3: 100135, 2023. 2

[4] James D. Crall, Nick Gravish, Andrew M. Mountcastle, and Stacey A. Combes. BEEtag: A Low-Cost, Image-Based Tracking System for the Study of Animal Behavior and Locomotion. PLoS ONE, 10(9):e0136487, 2015. 1, 2

[5] Adam L. Cronin and Michael P. Schwarz. Life Cycle and Social Behavior in a Heathland Population of Exoneura robusta (Hymenoptera: Apidae): Habitat Influences Opportunities for Sib Rearing in a Primitively Social Bee. Annals of the Entomological Society of America, 92(5):707–716, 1999. 2

[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3

[7] R. L. Dennis, R. C. Newberry, H.-W. Cheng, and I. Estevez. Appearance Matters: Artificial Marking Alters Aggression and Stress. Poultry Science, 87(10):1939–1946, 2008. 1

[8] Hiruni Samadi Galpayage Dona, Cwyn Solvi, Amelia Kowalewska, Kaarle Makel ¨ a, HaDi MaBouDi, and Lars ¨ Chittka. Do bumble bees play? Animal Behaviour, 194: 239–251, 2022. 2

[9] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc¸ois Laviolette, Mario March, and Victor Lempitsky. Domain-Adversarial Training of Neural Networks. Journal of Machine Learning Research, 17(59):1–35, 2016. 4

[10] Linden Gillbank. Hawkweeds: a recent discovery in Victoria’s Alps and a taxonomic name change, 2013. Accessed: 2024-05-22, https://invasives.org.au/blog/hawkweeds-arecent-discovery-in-victorias-alps-and-a-taxonomic-namechange/. 4

[11] Gerald Kastberger, Sarah Radloff, and Gerhard Kranner. Individuality of wing patterning in Giant honey bees (Apis laboriosa). Apidologie, 34(3):311–318, 2003. 2

[12] Shengcai Liao, Yang Hu, Xiangyu Zhu, and Stan Z. Li. Person Re-Identification by Local Maximal Occurrence Representation and Metric Learning. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2197–2206, 2015. 2

[13] Giuseppe Lisanti, Iacopo Masi, Andrew D. Bagdanov, and Alberto Del Bimbo. Person Re-Identification by Iterative ReWeighted Sparse Ranking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(8):1629–1642, 2015. 2

[14] Danielle P. Mersch, Alessandro Crespi, and Laurent Keller. Tracking Individuals Shows Spatial Fidelity Is a Key Regulator of Ant Social Organization. Science, 340(6136):1090– 1093, 2013. 1, 2

[15] Luke Meyers, Josue Rodr ´ ´ıguez Cordero, Carlos Corrada Bravo, Fanfan Noel, Jose Agosto-Rivera, Tugrul Giray, and ´ Remi M ´ egret. Towards Automatic Honey Bee Flower-Patch ´ Assays with Paint Marking Re-Identification. arXiv, 2023. 1, 2

[16] Nihal Murali, Jonathan Schneider, Joel D. Levine, and Graham W. Taylor. Classification and Re-Identification of Fruit Fly Individuals Across Days with Convolutional Neural Networks. 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 00:570–578, 2019. 1, 2, 4

[17] Thi Nha Ngo, Dan Jeric Arcega Rustia, En-Cheng Yang, and Ta-Te Lin. Automated monitoring and analyses of honey bee pollen foraging behavior using a deep learning-based imaging system. Computers and Electronics in Agriculture, 187: 106239, 2021. 4

[18] Stephen M. Pizer, E. Philip Amburn, John D. Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romeny, John B. Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing, 39(3):355–368, 1987. 3

[19] Florian Schroff, Dmitry Kalenictbhenko, and J.ames Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815–823, 2015. 2

[20] Yaniv Taigman, Ming Yang, Marc’ Aurelio Ranzato, and Lior Wolf. DeepFace: Closing the Gap to Human-Level Performance in Face Verification. 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 1701– 1708, 2014. 2

[21] Frederic Tausch, Simon Stock, Julian Fricke, and Olaf Klein. Bumblebee Re-Identification Dataset. 2020 IEEE Winter Applications of Computer Vision Workshops (WACVW), 00:35– 37, 2020. 1

[22] Fernando Wario, Benjamin Wild, Margaret J. Couvillon, Raul Rojas, and Tim Landgraf. Automatic methods for long-term tracking and the detection and decoding of communication dances in honeybees. Frontiers in Ecology and Evolution, 3:103, 2015. 4

[23] Fernando Wario, Benjamin Wild, Raul Rojas, and Tim Landgraf. Automatic detection and decoding of honey bee waggle dances. PLOS ONE, 12(12):e0188626, 2017. 4

[24] Shangxuan Wu, Ying-Cong Chen, Xiang Li, An-Cong Wu, Jin-Jie You, and Wei-Shi Zheng. An Enhanced Deep Feature Representation for Person Re-Identification. 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1–8, 2016. 2