JZUS - Journal of Zhejiang University SCIENCE

Frontiers of Information Technology & Electronic Engineering

Accepted manuscript available online (unedited version)

Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting

Author(s): Jiaqi GAO, Jingqi LI, Hongming SHAN, Yanyun QU, James Z. WANG, Fei-Yue WANG, Junping ZHANG
Affiliation(s): Shanghai Key Laboratory of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai 200433, China; more
Corresponding email(s): jqgao20@fudan.edu.cn, feiyue.wang@ia.ac.cn, jpzhang@fudan.edu.cn
Key Words: Crowd counting; Knowledge distillation; Lifelong learning

Share this article to： More \|Next Paper >>>

Jiaqi GAO, Jingqi LI, Hongming SHAN, Yanyun QU, James Z. WANG, Fei-Yue WANG, Junping ZHANG. Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting[J]. Frontiers of Information Technology & Electronic Engineering , 2023, 24(1): 187-202.

@article{title="Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting",
author="Jiaqi GAO, Jingqi LI, Hongming SHAN, Yanyun QU, James Z. WANG, Fei-Yue WANG, Junping ZHANG",
journal="Frontiers of Information Technology & Electronic Engineering",
volume="24",
number="2",
pages="187-202",
year="2023",
publisher="Zhejiang University Press & Springer",
doi="10.1631/FITEE.2200380"
}

%0 Journal Article
%T Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting
%A Jiaqi GAO
%A Jingqi LI
%A Hongming SHAN
%A Yanyun QU
%A James Z. WANG
%A Fei-Yue WANG
%A Junping ZHANG
%J Frontiers of Information Technology & Electronic Engineering
%V 24
%N 2
%P 187-202
%@ 1869-1951
%D 2023
%I Zhejiang University Press & Springer

TY - JOUR
T1 - Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting
A1 - Jiaqi GAO
A1 - Jingqi LI
A1 - Hongming SHAN
A1 - Yanyun QU
A1 - James Z. WANG
A1 - Fei-Yue WANG
A1 - Junping ZHANG
J0 - Frontiers of Information Technology & Electronic Engineering
VL - 24
IS - 2
SP - 187
EP - 202
%@ 1869-1951
Y1 - 2023
PB - Zhejiang University Press & Springer
ER -

Abstract
Chinese Summary
Academic Network
Reviewer Comment

Abstract: Crowd counting has important applications in public safety and pandemic control. A robust and practical crowd counting system has to be capable of continuously learning with the newly incoming domain data in real-world scenarios instead of fitting one domain only. Off-the-shelf methods have some drawbacks when handling multiple domains: (1) the models will achieve limited performance (even drop dramatically) among old domains after training images from new domains due to the discrepancies in intrinsic data distributions from various domains, which is called catastrophic forgetting; (2) the well-trained model in a specific domain achieves imperfect performance among other unseen domains because of domain shift; (3) it leads to linearly increasing storage overhead, either mixing all the data for training or simply training dozens of separate models for different domains when new ones are available. To overcome these issues, we investigate a new crowd counting task in incremental domain training setting called lifelong crowd counting. Its goal is to alleviate catastrophic forgetting and improve the generalization ability using a single model updated by the incremental domains. Specifically, we propose a self-distillation learning framework as a benchmark (forget less, count better, or FLCB) for lifelong crowd counting, which helps the model leverage previous meaningful knowledge in a sustainable manner for better crowd counting to mitigate the forgetting when new data arrive. A new quantitative metric, normalized Backward Transfer (nBwT), is developed to evaluate the forgetting degree of the model in the lifelong learning process. Extensive experimental results demonstrate the superiority of our proposed benchmark in achieving a low catastrophic forgetting degree and strong generalization ability.

忘得少，数得好：一种域增量式自蒸馏终身人群计数基准

高佳琪¹，李婧琦¹，单洪明^2,3，曲延云⁴，王则⁵，王飞跃⁶，张军平¹
¹复旦大学计算机科学技术学院上海市智能信息处理重点实验室，中国上海市，200433
²复旦大学类脑智能科学与技术研究院，中国上海市，200433
³上海脑科学与类脑研究中心，中国上海市，201210
⁴厦门大学信息科学与技术学院，中国厦门市，361005
⁵宾夕法尼亚州立大学信息科学与技术学院，美国宾夕法尼亚州，16802
⁶中国科学院自动化研究所复杂系统管理与控制国家重点实验室，中国北京市，100190
摘要：人群计数在公共安全和流行病控制方面具有重要应用。一个鲁棒且实用的人群计数系统须能够在真实场景中不断学习持续到来的新域数据，而非仅仅拟合某一单域的数据分布。现有方法在处理多个域的数据时有一些不足之处：（1）由于来自不同域的固有数据分布之间的差异，模型在训练来自新域的图像数据后在旧域中的性能可能会变得十分有限（甚至急剧下降），这种现象被称为灾难性遗忘；（2）由于域分布的偏移，在某一特定域数据中训练好的模型在其他未见域中通常表现不佳；（3）处理多个域的数据通常会导致存储开销的线性增长，例如混合来自所有域的数据进行训练，或者是简单地为每一个域的数据单独训练一个模型。为克服这些问题，我们探索了在域增量式训练设置下一种新的人群计数任务，即终身人群计数。它的目标是通过使用单个模型持续不断地学习新域数据以减轻灾难性遗忘并提高泛化能力。具体来说，提出一种自蒸馏学习框架作为终身人群计数的基准模型（forget less，count better，FLCB），这有助于模型可持续地利用之前学到的有意义的知识来更好地对人数进行估计，以减少训练新数据后对旧数据的遗忘。此外，设计了一种新的定量评价指标，即归一化后向迁移（normalized Backward Transfer，nBwT），用于评估模型在终身学习过程中的遗忘程度。大量实验结果证明了该模型的优越性，即较低的灾难性遗忘度和较强的泛化能力。

关键词组：人群计数；知识蒸馏；终身学习

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference

[1]Bai S, He ZQ, Qiao Y, et al., 2020. Adaptive dilated network with self-correction supervision for counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.4594-4603.

[2]Belouadah E, Popescu A, 2019. IL2M: class incremental learning with dual memory. Proc IEEE/CVF Int Conf on Computer Vision, p.583-592.

[3]Boominathan L, Kruthiventi SSS, Babu RV, 2016. CrowdNet: a deep convolutional network for dense crowd counting. Proc 24^th ACM Int Conf on Multimedia, p.640-644.

[4]Cao XK, Wang ZP, Zhao YY, et al., 2018. Scale aggregation network for accurate and efficient crowd counting. Proc 15^th European Conf on Computer Vision, p.734-750.

[5]Caron M, Misra I, Mairal J, et al., 2020. Unsupervised learning of visual features by contrasting cluster assignments. Proc 34^th Int Conf on Neural Information Processing Systems, p.9912-9924.

[6]Chan AB, Vasconcelos N, 2009. Bayesian Poisson regression for crowd counting. Proc 12^th IEEE Int Conf on Computer Vision, p.545-551.

[7]Chen BH, Yan ZY, Li K, et al., 2021. Variational attention: propagating domain-specific knowledge for multi-domain learning in crowd counting. Proc IEEE/CVF Int Conf on Computer Vision, p.16065-16075.

[8]Chen T, Kornblith S, Norouzi M, et al., 2020. A simple framework for contrastive learning of visual representations. Proc 37^th Int Conf on Machine Learning, p.1597-1607.

[9]Chen XY, Bin YR, Sang N, et al., 2019. Scale pyramid network for crowd counting. Proc IEEE Winter Conf on Applications of Computer Vision, p.1941-1950.

[10]Dalal N, Triggs B, 2005. Histograms of oriented gradients for human detection. Proc IEEE Computer Society Conf on Computer Vision and Pattern Recognition, p.886-893.

[11]Dollar P, Wojek C, Schiele B, et al., 2012. Pedestrian detection: an evaluation of the state of the art. IEEE Trans Patt Anal Mach Intell, 34(4):743-761.

[12]Grill JB, Strub F, Altché F, et al., 2020. Bootstrap your own latent a new approach to self-supervised learning. Proc 34^th Int Conf on Neural Information Processing Systems, p.21271-21284.

[13]Guo D, Li K, Zha ZJ, et al., 2019. DADNet: dilated-attention-deformable ConvNet for crowd counting. Proc 27^th ACM Int Conf on Multimedia, p.1823-1832.

[14]Han T, Gao JY, Yuan Y, et al., 2020. Focus on semantic consistency for cross-domain crowd understanding. Proc IEEE Int Conf on Acoustics, Speech and Signal Processing, p.1848-1852.

[15]He KM, Fan HQ, Wu YX, et al., 2020. Momentum contrast for unsupervised visual representation learning. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.9729-9738.

[16]He YJ, Sick B, 2021. CLeaR: an adaptive continual learning framework for regression tasks. AI Persp, 3(1):2.

[17]Huang ZZ, Chen J, Zhang JP, et al., 2022. Learning representation for clustering via prototype scattering and positive sampling. IEEE Trans Patt Anal Mach Intell, early access.

[18]Idrees H, Tayyab M, Athrey K, et al., 2018. Composition loss for counting, density map estimation and localization in dense crowds. Proc 15^th European Conf on Computer Vision, p.532-546.

[19]Jiang SQ, Lu XB, Lei YJ, et al., 2020. Mask-aware networks for crowd counting. IEEE Trans Circ Syst Video Technol, 30(9):3119-3129.

[20]Jiang XH, Zhang L, Xu ML, et al., 2020a. Attention scaling for crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.4706-4715.

[21]Jiang XH, Zhang L, Lv P, et al., 2020b. Learning multi-level density maps for crowd counting. IEEE Trans Neur Netw Learn Syst, 31(8):2705-2715.

[22]Kirkpatrick J, Pascanu R, Rabinowitz N, et al., 2017. Overcoming catastrophic forgetting in neural networks. PNAS, 114(13):3521-3526.

[23]Leibe B, Seemann E, Schiele B, 2005. Pedestrian detection in crowded scenes. Proc IEEE/CVF Computer Society Conf on Computer Vision and Pattern Recognition, p.878-885.

[24]Li YH, Zhang XF, Chen DM, 2018. CSRNet: dilated convolutional neural networks for understanding the highly congested scenes. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.1091-1100.

[25]Li ZZ, Hoiem D, 2018. Learning without forgetting. IEEE Trans Patt Anal Mach Intell, 40(12):2935-2947.

[26]Liu L, Lu H, Xiong HP, et al., 2020. Counting objects by blockwise classification. IEEE Trans Circ Syst Video Technol, 30(10):3513-3527.

[27]Liu LB, Qiu ZL, Li GB, et al., 2019. Crowd counting with deep structured scale integration network. Proc IEEE/CVF Int Conf on Computer Vision, p.1774-1783.

[28]Liu LB, Chen JQ, Wu HF, et al., 2021. Cross-modal collaborative representation learning and a large-scale RGBT benchmark for crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.4823-4833.

[29]Liu N, Long YC, Zou CQ, et al., 2019. ADCrowdNet: an attention-injective deformable convolutional network for crowd understanding. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.3225-3234.

[30]Liu WZ, Salzmann M, Fua P, 2019. Context-aware crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.5099-5108.

[31]Liu WZ, Durasov N, Fua P, 2022. Leveraging self-supervision for cross-domain crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.5341-5352.

[32]Lopez-Paz D, Ranzato M, 2017. Gradient episodic memory for continual learning. Proc 31^st Int Conf on Neural Information Processing Systems, p.6467-6476.

[33]Lowe DG, 1999. Object recognition from local scale-invariant features. Proc 7^th IEEE Int Conf on Computer Vision, p.1150-1157.

[34]Luo A, Yang F, Li X, et al., 2020. Hybrid graph neural networks for crowd counting. Proc 34^th AAAI Conf on Artificial Intelligence, p.11693-11700.

[35]Ma ZH, Wei X, Hong XP, et al., 2019. Bayesian loss for crowd count estimation with point supervision. Proc IEEE/CVF Int Conf on Computer Vision, p.6142-6151.

[36]Ma ZH, Wei X, Hong XP, et al., 2020. Learning scales from points: a scale-aware probabilistic model for crowd counting. Proc 28^th ACM Int Conf on Multimedia, p.220-228.

[37]Ma ZH, Hong XP, Wei X, et al., 2021. Towards a universal model for cross-dataset crowd counting. Proc IEEE/CVF Int Conf on Computer Vision, p.3205-3214.

[38]Niu C, Wang G, 2022a. Self-supervised representation learning with MUlti-Segmental Informational Coding (MUSIC). https://arxiv.org/abs/2206.06461

[39]Niu C, Wang G, 2022b. Unsupervised contrastive learning based transformer for lung nodule detection. Phys Med Biol, 67(20):204001.

[40]Niu C, Li MZ, Fan FL, et al., 2020. Suppression of correlated noise with similarity-based unsupervised deep learning. https://arxiv.org/abs/2011.03384

[41]Niu C, Shan HM, Wang G, 2022. SPICE: semantic pseudo-labeling for image clustering. IEEE Trans Image Process, 31:7264-7278.

[42]Rebuffi SA, Kolesnikov A, Sperl G, et al., 2017. iCaRL: incremental classifier and representation learning. Proc IEEE Conf on Computer Vision and Pattern Recognition, p.2001-2010.

[43]Rusu AA, Rabinowitz NC, Desjardins G, et al., 2016. Progressive neural networks. https://arxiv.org/abs/1606.04671

[44]Sam DB, Surya S, Babu RV, 2017. Switching convolutional neural network for crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.5744-5752.

[45]Shi ZL, Mettes P, Snoek C, 2019. Counting with focus for free. Proc IEEE/CVF Int Conf on Computer Vision, p.4200-4209.

[46]Sindagi VA, Patel VM, 2017. Generating high-quality crowd density maps using contextual pyramid CNNs. Proc IEEE Int Conf on Computer Vision, p.1861-1870.

[47]Sindagi VA, Patel VM, 2020. HA-CCN: hierarchical attention-based crowd counting network. IEEE Trans Image Process, 29:323-335.

[48]Sindagi V, Yasarla R, Patel V, 2019. Pushing the frontiers of unconstrained crowd counting: new dataset and benchmark method. Proc IEEE/CVF Int Conf on Computer Vision, p.1221-1231.

[49]Song QY, Wang CA, Wang YB, et al., 2021. To choose or to fuse? Scale selection for crowd counting. Proc 35^th AAAI Conf on Artificial Intelligence, p.2576-2583.

[50]Tan X, Tao C, Ren TW, et al., 2019. Crowd counting via multi-layer regression. Proc 27^th ACM Int Conf on Multimedia, p.1907-1915.

[51]Tian YK, Lei YM, Zhang JP, et al., 2020. PaDNet: pan-density crowd counting. IEEE Trans Image Process, 29:2714-2727.

[52]Tuzel O, Porikli F, Meer P, 2008. Pedestrian detection via classification on Riemannian manifolds. IEEE Trans Patt Anal Mach Intell, 30(10):1713-1727.

[53]Wang BY, Liu HD, Samaras D, et al., 2020. Distribution matching for crowd counting. Proc 34^th Int Conf on Neural Information Processing Systems, p.1595-1607.

[54]Wang C, Zhang H, Yang L, et al., 2015. Deep people counting in extremely dense crowds. Proc 23^rd ACM Int Conf on Multimedia, p.1299-1302.

[55]Wang Q, Gao JY, Lin W, et al., 2019. Learning from synthetic data for crowd counting in the wild. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.8198-8207.

[56]Wang Q, Gao JY, Lin W, et al., 2021. NWPU-crowd: a large-scale benchmark for crowd counting and localization. IEEE Trans Patt Anal Mach Intell, 43(6):2141-2149.

[57]Wang Q, Han T, Gao JY, et al., 2022. Neuron linear transformation: modeling the domain shift for crowd counting. IEEE Trans Neur Netw Learn Syst, 33(8):3238-3250.

[58]Wu QQ, Wan J, Chan AB, 2021. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. Proc 29^th ACM Int Conf on Multimedia, p.658-666.

[59]Xiong HP, Lu H, Liu CX, et al., 2019. From open set to closed set: counting objects by spatial divide-and-conquer. Proc IEEE/CVF Int Conf on Computer Vision, p.8362-8371.

[60]Yan ZY, Li PY, Wang B, et al., 2021. Towards learning multi-domain crowd counting. IEEE Trans Circ Syst Video Technol, early access.

[61]Yang YF, Li GR, Wu Z, et al., 2020. Reverse perspective network for perspective-aware object counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.4374-4383.

[62]Zhang C, Li HS, Wang XG, et al., 2015. Cross-scene crowd counting via deep convolutional neural networks. Proc IEEE Conf on Computer Vision and Pattern Recognition, p.833-841.

[63]Zhang Q, Lin W, Chan AB, 2021. Cross-view cross-scene multi-view crowd counting. Proc IEEE/CVF Conf on Computer Vision and Pattern Recognition, p.557-567.

[64]Zhang YY, Zhou DS, Chen SQ, et al., 2016. Single-image crowd counting via multi-column convolutional neural network. Proc IEEE Conf on Computer Vision and Pattern Recognition, p.589-597.

[65]Zhao MM, Zhang CY, Zhang J, et al., 2020. Scale-aware crowd counting via depth-embedded convolutional neural networks. IEEE Trans Circ Syst Video Technol, 30(10):3651-3662.

[66]Zhu JY, Park T, Isola P, et al., 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. Proc IEEE Int Conf on Computer Vision, p.2223-2232.

[67]Zhu L, Zhao ZJ, Lu C, et al., 2019. Dual path multi-scale fusion networks with attention for crowd counting. https://arxiv.org/abs/1902.01115

[68]Zou ZK, Qu XY, Zhou P, et al., 2021. Coarse to fine: domain adaptive crowd counting via adversarial scoring network. Proc 29^th ACM Int Conf on Multimedia, p.2185-2194.

Open peer comments: Debate/Discuss/Question/Opinion

<1>

- Go to

忘得少，数得好：一种域增量式自蒸馏终身人群计数基准

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference