Interpretability methods often measure the contribution of an input feature to an image classifier's decisions by heuristically removing it via e.g. blurring, adding noise, or graying out, which often produce unrealistic, out-of-samples. Instead, we propose to integrate a generative inpainter into three representative attribution methods to remove an input feature. Compared to the original counterparts, our methods (1) generate more plausible counterfactual samples under the true data generating process; (2) are more robust to hyperparameter settings; and (3) localize objects more accurately. Our findings were consistent across both ImageNet and Places365 datasets and two different pairs of classifiers and inpainters.
Fiducial markers have been playing an important role in augmented reality (AR), robot navigation, and general applications where the relative pose between a camera and an object is required. We introduce TopoTag, a robust and scalable topological fiducial marker system, which supports reliable and accurate pose estimation from a single image. TopoTag uses topological and geometrical information in marker detection to achieve higher robustness. Without sacrificing bits for higher recall and precision like previous systems, TopoTag can use full bits for ID encoding and supports tens of thousands unique IDs and easily extends to millions and more by adding more bits, thus achieves perfect scalability. We collect a large dataset including in total 169,713 images for evaluation, involving in-plane and out-of-plane rotation, image blur, different distances and various backgrounds, etc. Experiments show that TopoTag significantly outperforms previous fiducial marker systems in terms of various metrics, including detection accuracy, vertex jitter, pose jitter and accuracy, etc. In addition, TopoTag supports occlusion as long as main tag topological structure is maintained and flexible shape design where users can customize inter and outer marker shapes. Our dataset, marker design and detection algorithm are public to the community.
Variational method and deep learning method are two mainstream powerful approaches to solve inverse problems in computer vision. To take advantages of advanced optimization algorithms and powerful representation ability of deep neural networks, we propose a novel deep network for image reconstruction. The architecture of this network is inspired by our proposed accelerated extra proximal gradient algorithm. It is able to incorporate non-local operation to exploit the non-local self-similarity of the images and to learn the nonlinear transform, under which the solution is sparse. All the parameters in our network are learned from minimizing a loss function. Our experimental results show that our network outperforms several state-of-the-art deep networks with almost the same number of learnable parameter.
We present a context aware object detection method based on a retrieve-and-transform scene layout model. Given an input image, our approach first retrieves a coarse scene layout from a codebook of typical layout templates. In order to handle large layout variations, we use a variant of the spatial transformer network to transform and refine the retrieved layout, resulting in a set of interpretable and semantically meaningful feature maps of object locations and scales. The above steps are implemented as a Layout Transfer Network which we integrate into Faster RCNN to allow for joint reasoning of object detection and scene layout estimation. Extensive experiments on three public datasets verified that our approach provides consistent performance improvements to the state-of-the-art object detection baselines on a variety of challenging tasks in the traffic surveillance and the autonomous driving domains.
Although many studies have been made on markerless motion capture, it has not been applied to real sports or concerts. In this paper, we propose a markerless motion capture method with spatiotemporal accuracy and smoothness from multiple cameras, even in wide and multi-person environments. The key idea is predicting each person's 3D pose and determining the bounding box of multi-camera images small enough. This prediction and spatiotemporal filtering based on human skeletal structure eases 3D reconstruction of the person and yields accuracy. The accurate 3D reconstruction is then used to predict the bounding box of each camera image in the next frame. This is a feedback from 3D motion to 2D pose, and provides a synergetic effect to the total performance of video motion capture. We demonstrate the method using various datasets and a real sports field. The experimental results show the mean per joint position error was 31.6mm and the percentage of correct parts was 99.3% under five people moving dynamically, with satisfying the range of motion. Video demonstration, datasets, and additional materials are posted on our project page.
Semantic segmentation (i.e. image parsing) aims to annotate each image pixel with its corresponding semantic class label. Spatially consistent labeling of the image requires an accurate description and modeling of the local contextual information. Segmentation result is typically improved by Markov Random Field (MRF) optimization on the initial labels. However this improvement is limited by the accuracy of initial result and how the contextual neighborhood is defined. In this paper, we develop generalized and flexible contextual models for segmentation neighborhoods in order to improve parsing accuracy. Instead of using a fixed segmentation and neighborhood definition, we explore various contextual models for fusion of complementary information available in alternative segmentations of the same image. In other words, we propose a novel MRF framework that describes and optimizes the contextual dependencies between multiple segmentations. Simulation results on two common datasets demonstrate significant improvement in parsing accuracy over the baseline approaches.
The analysis of the tumor environment on digital histopathology slides is becoming key for the understanding of the immune response against cancer, supporting the development of novel immuno-therapies. We introduce here a novel deep learning solution to the related problem of tumor epithelium segmentation. While most existing deep learning segmentation approaches are trained on time-consuming and costly manual annotation on single stain domain (PD-L1), we leverage here semi-automatically labeled images from a second stain domain (Cytokeratin-CK). We introduce an end-to-end trainable network that jointly segment tumor epithelium on PD-L1 while leveraging unpaired image-to-image translation between CK and PD-L1, therefore completely bypassing the need for serial sections or re-staining of slides. Extending the method to differentiate between PD-L1 positive and negative tumor epithelium regions enables the automated estimation of the PD-L1 Tumor Cell (TC) score. Quantitative experimental results demonstrate the accuracy of our approach against state-of-the-art segmentation methods.
This paper addresses the construction of a short-vector (128D) image representation for large-scale image and particular object retrieval. In particular, the method of joint dimensionality reduction of multiple vocabularies is considered. We study a variety of vocabulary generation techniques: different k-means initializations, different descriptor transformations, different measurement regions for descriptor extraction. Our extensive evaluation shows that different combinations of vocabularies, each partitioning the descriptor space in a different yet complementary manner, results in a significant performance improvement, which exceeds the state-of-the-art.
In this work we target the problem of hate speech detection in multimodal publications formed by a text and an image. We gather and annotate a large scale dataset from Twitter, MMHS150K, and propose different models that jointly analyze textual and visual information for hate speech detection, comparing them with unimodal detection. We provide quantitative and qualitative results and analyze the challenges of the proposed task. We find that, even though images are useful for the hate speech detection task, current multimodal models cannot outperform models analyzing only text. We discuss why and open the field and the dataset for further research.
Quantization plays an important role for energy-efficient deployment of deep neural networks (DNNs) on resource-limited devices. Post-training quantization is crucial since it does not require retraining or accessibility to the full training dataset. The conventional post-training uniform quantization scheme achieves satisfactory results by converting DNNs from full-precision to 8-bit integers, however, it suffers from significant performance degradation when quantizing to lower precision such as 4 bits. In this paper, we propose a piecewise linear quantization method to enable accurate post-training quantization. Inspired from the fact that the weight tensors have bell-shaped distributions with long tails, our approach breaks the entire quantization range into two non-overlapping regions for each tensor, with each region being assigned an equal number of quantization levels. The optimal break-point that divides the entire range is found by minimizing the quantization error. Extensive results show that the proposed method achieves state-of-the-art performance on image classification, semantic segmentation and object detection. It is possible to quantize weights to 4 bits without retraining while nearly maintaining the performance of the original full-precision model.