Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection

By Grant Van Horn, Steve Branson, Ryan Farrell, S. Haber, Jessie Barry, Panagiotis G. Ipeirotis, Pietro Perona, Serge J. Belongie. 2015

720 citations recorded by the source corpus.

Subfields: Computer Science Applications, Computer Vision and Pattern Recognition, Developmental Biology, Ecological Modeling

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Science Applications

In Computer Vision and Pattern Recognition

In Developmental Biology

In Ecological Modeling

Citations in the corpus

Fixing the train-test resolution discrepancy (2019, methodology): “We evaluated it on the following datasets: iNaturalist 2017 [16], Stanford Cars [19], CUB-200-2011 [40], Oxford 102 Flowers [26], Oxford-IIIT Pets [28], NABirds [39] and Birdsnap [4].”
Visual Prompt Tuning (2022, methodology): “0 } [72] Fine-grained bird species recognition 55 21,536 ⋆ 2,393 ⋆ 24,633 Oxford Flowers [59] Fine-grained flower species recognition 102 1,020 1,020 6,149 Stanford Dogs [41] Fine-grained dog species recognition 120 10,800 ⋆ 1,200 ⋆ 8,580 Stanford Cars [23] Fine-grained Datasets and pre-trained…”
NDD20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation (2020, background): “Datasets often focus on a particular subset of animals such as pets commonly found in homes (9, 12) or birds (14, 17).”
The Unreasonable Effectiveness of Noisy Data for Fine-Grained Recognition (2015, methodology): “[52] further measure all sources of noise combined in Flickr to be only 16% when searching for bird species.”
Recognition in Terra Incognita (2018, background): “Natural world datasets such as the iNaturalist dataset [1], CUB200 [32], Oxford Flowers [33], LeafSnap [34], and NABirds700 [35] are focused on finegrained species classification and detection.”
Bilinear CNN Models for Fine-Grained Visual Recognition (2015, background): “4% label noise hence some of these errors may be incorrect [34].”
Neural Activation Constellations: Unsupervised Part Model Discovery with Convolutional Networks (2015, neutral): “For fine-grained classification, we evaluate our approach on CUB200-2011 [35] (200 classes, 11788 images), NA birds [34] (555 classes, 48562 images), Stanford dogs [20] (120 classes, 20580 images), Oxford flowers 102 [24] (102 classes, 8189 images), and Oxford-IIIT Pets [25] (37 classes, 7349 images”
Deep Layer Aggregation (2017, methodology): “(Wah et al. 2011), Car (Krause et al. 2013), Plane (Maji et al. 2013), Food (Bossard et al. 2014), and NABirds (Van Horn et al. 2015).”
The Devil is in the Tails: Fine-grained Classification in the Wild (2017, background): “Some of the recent ﬁne-grained papers have investigated augmenting the original datasets with additional data from the web [26, 50, 58, 59].”
The iNaturalist Challenge 2017 Dataset (2017, background): “Even manually vetted datasets such as ImageNet [31] have been reported to contain up to four percent error for some fine-grained categories [38].”
Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning (2018, background): “Examples include recognizing natural categories such as species of birds [58, 54], dogs [28] and plants [39, 59]; or man-made categories such as car make & model [32, 63].”
Few-Shot Unsupervised Image-to-Image Translation (2019, neutral): “• Birds [44].”
Bilinear CNNs for Fine-grained Visual Recognition (2015, methodology): “[64] obtains 75% accuracy using AlexNet and part annotations at test time, while the “neural activation constellations” approach [57] obtains 76.”
DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort (2021, methodology): “For bird, we train our own StyleGAN models on NABirds [50], which contains 48k images.”
Improving Visual Prompt Tuning for Self-supervised Vision Transformers (2023, methodology): “On the FGVC benchmark, which encompasses five fine-grained classification tasks (Wah et al., 2011; Van Horn et al., 2015; Khosla et al., 2011; Gebru et al., 2017; Nilsback & Zisserman, 2008), we achieve an average accuracy of 73.4% for MAE and 83.0% for MoCo v3.”
Bugs in the Data: How ImageNet Misrepresents Biodiversity (2022, positive): “To date, these studies have focused on the the “people” and “object” classes of ImageNet, with the notable exception of a 2015 study that looked at the bird classes in ImageNet, hypothesizing class error rates of at least 4% (Van Horn et al. 2015).”
The iNaturalist Species Classification and Detection Dataset (2017, background): “Even manually vetted datasets such as ImageNet [32] have been reported to contain up to 4% error for some fine-grained categories [40].”
Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning (2022, methodology): “All experiments are implemented with pre-trained ViT-B/16 models on CIFAR-100 and the results are shown in Table 6.”