CORe50: a New Dataset and Benchmark for Continuous Object Recognition

By Vincenzo Lomonaco, Davide Maltoni. 2017

584 citations recorded by the source corpus.

Continuous/Lifelong learning of high-dimensional data streams is a challenging research problem. In fact, fully retraining models each time new data become available is infeasible, due to computational and storage issues, while naive incremental strategies have been shown to suffer from catastrophic forgetting. In the context of real-world object recognition applications (e.g., robotic vision), where continuous learning is crucial, very few datasets and benchmarks are available to evaluate and compare emerging techniques. In this work we propose a new dataset and benchmark CORe50, specifically designed for continuous object recognition, and introduce baseline approaches for different continuous learning scenarios.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Evaluation & Benchmarks

How researchers see this paper

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

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

Citations in the corpus

ORBIT: A Real-World Few-Shot Dataset for Teachable Object Recognition (2021, methodology): “Core50 [27] is a video dataset captured on mobile phones for a continual learning recognition task.”
A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “Some recent work has focused on combinatorial scenarios [18], [40], [41], [42], [43] in order to simulate real-world complexity.”
Continuous Learning in Single-Incremental-Task Scenarios (2018, background): “Implementing step 2 as above proved to be suboptimal with respect to CWR approach (see Figure 8 of [18] for a comparison) where old class weights are re-initialized at each batch.”
Lifelong Learning of Spatiotemporal Representations With Dual-Memory Recurrent Self-Organization (2018, methodology): “…as new batches become available, (ii) a proposed CopyWeights with Re-init (CWR) method that skips layers fc6 and fc7 of the CNN (for details, see Lomonaco and Maltoni, 2017; page 7), and (iii) a cumulative approach where the learning is carried out by considering the current batch and all the…”
Memory Efficient Experience Replay for Streaming Learning (2018, methodology): “2) CORe50: CORe50 [52] is similar to iCub1, and both have 10 classes.”
Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, methodology): “For CORe50, we use a SGD optimizer of learning rate 0.003, and train all approaches with batch size of 32 for 40 epochs.”
First Session Adaptation: A Strong Replay-Free Baseline for Class-Incremental Learning (2023, methodology): “The datasets are CIFAR100, SVHN, dSprites-location, FGVC-Aircraft, Cars, and Letters while also including 4 extra datasets: CUB200 [32], CORE50 [15], iNaturalist [30], and DomainNet [22].”
CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning (2022, background): “One group of approaches expands a model’s architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54].”
A Simple Baseline that Questions the Use of Pretrained-Models in Continual Learning (2022, methodology): “Table 4 shows the results of the continual learning in domain incremental setting on CoRe50[14].”
Few-Shot Continual Active Learning by a Robot (2022, methodology): “Regularization techniques (LWF and EWC) also suffer from catastrophic forgetting and produce the same accuracy as FT. CWR and iCaRL produce slightly better accuracy, however they also suffer from catastrophic forgetting.”
Repeated Augmented Rehearsal: A Simple but Strong Baseline for Online Continual Learning (2022, methodology): “Implementation: We use a reduced ResNet-18 for all datasets following Mai et al. [2021], Aljundi et al.”
S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for Domain Incremental Learning (2022, methodology): “In our experiments, the total class number C is 2 for CDDB [36], 50 for CORe50 [45], and 345 for DomainNet [52] respectively.”
Simpler is Better: off-the-shelf Continual Learning Through Pretrained Backbones (2022, methodology): “Memory KiB class Params Model CIFAR100 [10] CIFAR10 [10] Core50 [13] Oxford Flowers102 [18] Tiny ImgNet200 [23]”
General Incremental Learning with Domain-aware Categorical Representations (2022, methodology): “Existing works on the general incremental learning typically focuse on the online class incremental learning setting [3, 18], which has to sacrifice the performance due to its strict computation/memory constraints.”
Learning to Prompt for Continual Learning (2021, methodology): “where there is only a single test set available, we report the final test accuracy following the common protocol [30, 52].”
Online Continual Learning Via Candidates Voting (2021, methodology): “• An online sampler is designed to select exemplars from sequentially available data stream through dynamic mean update criteria and we further study exemplar augmentation in feature space to achieve improved performance • We conduct extensive experiments on benchmark datasets including CIFAR-10 [24”
Wanderlust: Online Continual Object Detection in the Real World (2021, background): “Moreover, existing benchmarks are mostly constructed using static images [9, 14, 22] or samples from short object-centric video snippets (∼15 seconds) [28].”