Fine-tuning from a collection of models pre-trained on different domains (a "model zoo") is emerging as a technique to improve test accuracy in the low-data regime. However, model selection, i.e. how to pre-select the right model to fine-tune from a model zoo without performing any training, remains an open topic. We use a linearized framework to approximate fine-tuning, and introduce two new baselines for model selection -- Label-Gradient and Label-Feature Correlation. Since all model selection algorithms in the literature have been tested on different use-cases and never compared directly, we introduce a new comprehensive benchmark for model selection comprising of: i) A model zoo of single and multi-domain models, and ii) Many target tasks. Our benchmark highlights accuracy gain with model zoo compared to fine-tuning Imagenet models. We show our model selection baseline can select optimal models to fine-tune in few selections and has the highest ranking correlation to fine-tuning accuracy compared to existing algorithms.
A deraining network may be interpreted as a condition generator. Image degradation generated by the deraining network can be attributed to defective embedding features that serve as conditions. Existing image deraining methods usually ignore uncertainty-caused model errors that lower embedding quality and embed low-quality features into the model directly. In contrast, we replace low-quality features by latent high-quality features. The spirit of closed-loop feedback in the automatic control field is borrowed to obtain latent high-quality features. A new method for error detection and feature compensation is proposed to address model errors. Extensive experiments on benchmark datasets as well as specific real datasets demonstrate the advantage of the proposed method over recent state-of-the-art methods.
We show a simple NMS-free, end-to-end object detection framework, of which the network is a minimal modification to a one-stage object detector such as the FCOS detection model [Tian et al. 2019]. We attain on par or even improved detection accuracy compared with the original one-stage detector. It performs detection at almost the same inference speed, while being even simpler in that now the post-processing NMS (non-maximum suppression) is eliminated during inference. If the network is capable of identifying only one positive sample for prediction for each ground-truth object instance in an image, then NMS would become unnecessary. This is made possible by attaching a compact PSS head for automatic selection of the single positive sample for each instance (see Fig. 1). As the learning objective involves both one-to-many and one-to-one label assignments, there is a conflict in the labels of some training examples, making the learning challenging. We show that by employing a stop-gradient operation, we can successfully tackle this issue and train the detector. On the COCO dataset, our simple design achieves superior performance compared to both the FCOS baseline detector with NMS post-processing and the recent end-to-end NMS-free detectors. Our extensive ablation studies justify the rationale of the design choices.
We extend the classical tracking-by-detection paradigm to this tracking-any-object task. Solid detection results are first extracted from TAO dataset. Some state-of-the-art techniques like \textbf{BA}lanced-\textbf{G}roup \textbf{S}oftmax (\textbf{BAGS}\cite{li2020overcoming}) and DetectoRS\cite{qiao2020detectors} are integrated during detection. Then we learned appearance features to represent any object by training feature learning networks. We ensemble several models for improving detection and feature representation. Simple linking strategies with most similar appearance features and tracklet-level post association module are finally applied to generate final tracking results. Our method is submitted as \textbf{AOA} on the challenge website.
Optimizers that further adjust the scale of gradient, such as Adam, Natural Gradient (NG), etc., despite widely concerned and used by the community, are often found poor generalization performance, compared with Stochastic Gradient Descent (SGD). They tend to converge excellently at the beginning of training but are weak at the end. An immediate idea is to complement the strengths of these algorithms with SGD. However, a truncated replacement of optimizer often leads to a crash of the update pattern, and new algorithms often require many iterations to stabilize their search direction. Driven by this idea and to address this problem, we design and present a regularized natural gradient optimization algorithm with look-ahead strategy, named asymptotic natural gradient (ANG). According to the total iteration step, ANG dynamic assembles NG and Euclidean gradient, and updates parameters along the new direction using the intensity of NG. Validation experiments on CIFAR10 and CIFAR100 data sets show that ANG can update smoothly and stably at the second-order speed, and achieve better generalization performance.
This paper presents our proposed methods for domain adaptive pedestrian re-identification (Re-ID) task in Visual Domain Adaptation Challenge (VisDA-2020). Considering the large gap between the source domain and target domain, we focused on solving two biases that influenced the performance on domain adaptive pedestrian Re-ID and proposed a two-stage training procedure. At the first stage, a baseline model is trained with images transferred from source domain to target domain and from single camera to multiple camera styles. Then we introduced a domain adaptation framework to train the model on source data and target data simultaneously. Different pseudo label generation strategies are adopted to continuously improve the discriminative ability of the model. Finally, with multiple models ensembled and additional post processing approaches adopted, our methods achieve 76.56% mAP and 84.25% rank-1 on the test set. Codes are available at https://github.com/vimar-gu/Bias-Eliminate-DA-ReID
Existing pixel-level adversarial attacks on neural networks may be deficient in real scenarios, since pixel-level changes on the data cannot be fully delivered to the neural network after camera capture and multiple image preprocessing steps. In contrast, in this paper, we argue from another perspective that gaussian blurring, a common technique of image preprocessing, can be aggressive itself in specific occasions, thus exposing the network to real-world adversarial attacks. We first propose an adversarial attack demo named peak suppression (PS) by suppressing the values of peak elements in the features of the data. Based on the blurring spirit of PS, we further apply gaussian blurring to the data, to investigate the potential influence and threats of gaussian blurring to performance of the network. Experiment results show that PS and well-designed gaussian blurring can form adversarial attacks that completely change classification results of a well-trained target network. With the strong physical significance and wide applications of gaussian blurring, the proposed approach will also be capable of conducting real world attacks.
Existing deep-learning based monocular orientation estimation algorithms faces the problem of confusion between the anterior and posterior parts of the objects, caused by the feature similarity of such parts in typical objects in traffic scenes such as cars and pedestrians. While difficult to solve, the problem may lead to serious orientation estimation errors, and pose threats to the upcoming decision making process of the ego vehicle, since the predicted tracks of objects may have directions opposite to ground truths. In this paper, we mitigate this problem by proposing a pretraining method. The method focuses on predicting the left/right semicircle in which the orientation of the object is located. The trained semicircle prediction model is then integrated into the orientation angle estimation model which predicts a value in range $[0, \pi]$. Experiment results show that the proposed semicircle prediction enhances the accuracy of orientation estimation, and mitigates the problem stated above. With the proposed method, a backbone achieves similar state-of-the-art orientation estimation performance to existing approaches with well-designed network structures.
Convolutional neural networks (CNN) have been more and more applied in mobile robotics such as intelligent vehicles. Security of CNNs in robotics applications is an important issue, for which potential adversarial attacks on CNNs are worth research. Pooling is a typical step of dimension reduction and information discarding in CNNs. Such information discarding may result in mis-deletion and mis-preservation of data features which largely influence the output of the network. This may aggravate the vulnerability of CNNs to adversarial attacks. In this paper, we conduct adversarial attacks on CNNs from the perspective of network structure by investigating and exploiting the vulnerability of pooling. First, a novel adversarial attack methodology named Strict Layer-Output Manipulation (SLOM) is proposed. Then an attack method based on Strict Pooling Manipulation (SPM) which is an instantiation of the SLOM spirit is designed to effectively realize both type I and type II adversarial attacks on a target CNN. Performances of attacks based on SPM at different depths are also investigated and compared. Moreover, performances of attack methods designed by instantiating the SLOM spirit with different operation layers of CNNs are compared. Experiment results reflect that pooling tends to be more vulnerable to adversarial attacks than other operations in CNNs.