Vincenzo Lomonaco, Davide Maltoni
University of Bologna
• Deep Learning holds state-of-the-art performances in many tasks
• Prediction models mainly trained with huge and fixed datasets.
• Higher and realistic time-scale where different data distributions
becomes available only during time
• No access to past perception data
• Imperative to build on top of previously learned knowledge and
skills.
• Ideal for Robotics and real-time learning from streaming
perception data.
• Continuous Learning needs the presence of multiple (temporal
coherent and unconstrained) views of the same objects taken in
different sessions.
# Images 164,866
Format RGB-D
Image size 350x350
128x128
# Categories 10
# Obj. x Cat. 5
# Sessions 11
# img. x Sess. ~300
# Outdoor Sess. 3
Acquisition Sett. Hand held
Categories
Objects
Approaches:
• Re-train the entire model from scratch as soon as a new batch of
data is available.
• Update the model with only the new batch of data.
𝐵𝑎𝑡𝑐ℎ0 𝐵𝑎𝑡𝑐ℎ1 𝐵𝑎𝑡𝑐ℎ 𝑛
Initial Batch Incremental Batches

. . .
New Instances (NI)
• New training patterns of the same classes.
New Classes (NC)
• New training patterns of different classes.
New Instances and Classes (NIC)
• New training patterns of known and new classes.
(NC)
(NI)
(NC)
(NIC)
Dataset, Benchmark, code and additional
information freely available at:
vlomonaco.github.io/core50
In the poster we also report
very recent results with:
1. Learning without Forgetting
(LwF)
2. ElasticWeight Consolidation
(EWC)

CORe50: a New Dataset and Benchmark for Continual Learning and Object Recognition - Slides

  • 1.
    Vincenzo Lomonaco, DavideMaltoni University of Bologna
  • 2.
    • Deep Learningholds state-of-the-art performances in many tasks • Prediction models mainly trained with huge and fixed datasets.
  • 3.
    • Higher andrealistic time-scale where different data distributions becomes available only during time • No access to past perception data • Imperative to build on top of previously learned knowledge and skills. • Ideal for Robotics and real-time learning from streaming perception data.
  • 4.
    • Continuous Learningneeds the presence of multiple (temporal coherent and unconstrained) views of the same objects taken in different sessions.
  • 5.
    # Images 164,866 FormatRGB-D Image size 350x350 128x128 # Categories 10 # Obj. x Cat. 5 # Sessions 11 # img. x Sess. ~300 # Outdoor Sess. 3 Acquisition Sett. Hand held Categories Objects
  • 6.
    Approaches: • Re-train theentire model from scratch as soon as a new batch of data is available. • Update the model with only the new batch of data. 𝐵𝑎𝑡𝑐ℎ0 𝐵𝑎𝑡𝑐ℎ1 𝐵𝑎𝑡𝑐ℎ 𝑛 Initial Batch Incremental Batches  . . .
  • 7.
    New Instances (NI) •New training patterns of the same classes. New Classes (NC) • New training patterns of different classes. New Instances and Classes (NIC) • New training patterns of known and new classes.
  • 8.
  • 9.
  • 10.
    Dataset, Benchmark, codeand additional information freely available at: vlomonaco.github.io/core50 In the poster we also report very recent results with: 1. Learning without Forgetting (LwF) 2. ElasticWeight Consolidation (EWC)