Learning Discriminative Embeddings for Object Recognition on-the-fly

Abstract

We address the problem of learning to recognize new objects on-the-fly efficiently. When using CNNs, a typical approach for learning new objects is by fine-tuning the model. However, this approach relies on the assumption that the original training set is available and requires high-end computational resources for training the ever-growing dataset efficiently, which can be unfeasible for robots with limited hardware. To overcome these limitations, we propose a new architecture that: 1) Instead of predicting labels, it learns to generate discriminative and separable embeddings of an object's viewpoints by using a Supervised Triplet Loss, which is easier to implement than current smart mining techniques and the trained model can be applied to unseen objects. 2) Infers an object's identity efficiently by utilizing a lightweight classifier in the features embedding space, this keeps the inference time in the order of milliseconds and can be retrained efficiently when new objects are learned. We evaluate our approach on four real-world images datasets used for Robotics and Computer Vision applications: Amazon Robotics Challenge 2017 by MIT-Princeton, T-LESS, ToyBoX, and CORe50 datasets. Code available at [1].Miguel Lagunes-Fortiz thanks the Mexican Council of Science and Technology (CONACyT) for sponsoring his studies with the scholarship number 686450

Centroids Triplet Network and Temporally-Consistent Embeddings for In-Situ Object Recognition

“…Our work builds on the findings from [5] for learning to recognize objects on-the-fly. However, to make the scalable real-time recognition system that we are after, we propose the following contributions.…”

“…While [3], [4] were designed for domain adaptation between in-vitro and in-situ, they empirically demonstrated that a ConvNet can be used to learn new objects without having to retrain the model by performing the k-nearest neighbors search in the features space, where data points from the same object are close to each other and separated otherwise. [5] builds on the same idea of utilizing a discriminative ConvNet and simplifies the model into a single branch that uses a combination of Softmax and Triplet Loss for achieving stateof-the-art accuracy for learning objects on-the-fly. We follow this research direction of utilizing discriminative networks for learning new objects without the need for retraining, and therefore, we focus this literature review on supervised approaches that learn discriminative features by utilizing regularizers in the loss function.…”

“…Regardless of the architecture design, it is now broadly studied that the commonly used combination of Cross-Entropy Loss and the Softmax function in the last fully connected layer, a.k.a. Softmax Loss, does not explicitly optimize the feature embedding to enforce higher similarity for intra-class samples and diversity for inter-class samples [5], [6], [7], [8], [9], [10], [11], [12], [13].…”

“…State-of-the-art Euclidean regularizers include Center Loss [9], where the ConvNet learns centers and the clustering of data points around those ones; Triplet Center Loss [14], originally proposed for 3D object retrieval, proposes the incorporation of the Triplet Loss with Center Loss in order to enforce inter-class separability of clusters; Similarly, [5] combines the Triplet Loss with the Softmax Loss where the features for each task are separated into two different heads, in order to improve the accuracy for classifying novel objects. In all these approaches, a hyper parameter is used to balance the supervision signals.…”

An Automated Method for Biometric Handwritten Signature Authentication Employing Neural Networks

“…To train the network, we used the triplet loss algorithm. It is a well-established method that finds its uses in many areas, including image retrieval, object recognition, and biometric verification [22,29,30]. This approach especially lends itself to the problem of biometric verification, as there exist a considerable amount of classes (each person could be considered a separate class), and their total number is not known at the training time.…”