This challenge intends to explore how to leverage the knowledge learned from previous tasks that could generalize to new task effectively, and also how to efficiently memorize of previously learned tasks. Making the robot behaves like the human with knowledge transfer, association, and combination capabilities.
To our best knowledge, the provided lifelong object recognition dataset is the 1st one that explicitly indicates the task difficulty under the incremental setting, which is able to foster the lifelong/continual/incremental learning in a supervised/semi-supervised manner. Different from previous instance/class-incremental task, the difficulty-incremental learning is to test the model’s capability over continuous learning when faced with multiple environmental factors, such as illumination, occlusion, camera-object distances/angles, clutter, and context information in both low and high dynamic scenes.
We have provided 3 pre-trained models over the provided datasets. The participants can use and think about how to optimize based on them.
The participants who achieve the high ranking results are encouraged to deliver an oral presentation in IROS 2019 competition section and have a official award from IROS onsite.
Data details are shown on our “lifelong robotic vision” website.
If you have any questions about the dataset, pre-trained modules, or any other competition problem please open issues on our github . You are welcome to join the online discussion with our organizers.
Yes. To obtain the consistent and comparable results, you need to train your model and design your own learning algorithm that can train sequentially over provided 9 batches.
For example, when you train the 1st batch of datasets, you can only have access to train/validation data of 1st batch, but need to test over 1-9 batch test datasets; next when you train the model over 2nd batch of datasets, you can only have access to train/validation data of 2nd batch, but you can also keep some of validation data from 1st batch for learning current model. Finally, you need to provided your model sequentially trained on 9 batches of dataset.
For example, when you train the 1st batch of datasets, you can only have access to train/validation data of 1st batch, but need to test over 1-9 batch test datasets; next when you train the model over 2nd batch of datasets, you can only have access to train/validation data of 2nd batch, but you can also keep some of validation data from 1st batch for learning current model. That means the validation dataset can be kept for the sequential learning task. Finally, you need to provide your model that have been sequentially trained over 9 batches of datasets.
This scenario is quite common when the system is deployed as the real-world application, which should be able to update the model day after day, also the system memory is also valuable that we can not keep old dataset on the current systems, sometimes, we only select some coresets of the dataset that we have used. It is non-trivial to pick out the data that can be the summarizations of the encountered data. We constrain our learning/training procedure that approach this kind of real-world scenario.
Yes, you can submit the provided baseline models, but as we have tested, the results are not state-of-the-arts. The models provided can be used for your quicker familiar with the procedure.
Yes, we recommend you go through the recent review paper over this topic:Continual Lifelong Learning with Neural Networks: A Review.
OpenLORIS-object dataset is maintained by Qi She
This work is licensed under a Creative Commons Attribution 4.0 International License.