ICCV 2021 Papers

Model quantization has emerged as a mandatory technique for efficient inference with advanced Deep Neural Networks (DNN). It converts the model parameters in full precision (32-bit floating point) to the hardware friendly data representation with shorter bit-width, to not only reduce the model size but also simplify the computation complexity. Nevertheless, prior model quantization either suffers from the inefficient data encoding method thus leading to noncompetitive model compression rate, or requires time-consuming quantization aware training process. In this work, we propose a novel Adaptive Floating-Point (AFP) as a variant of standard IEEE-754 floating-point format, with flexible configuration of exponent and mantissa segments. Leveraging the AFP for model quantization (i.e., encoding the parameter) could significantly enhance the model compression rate without accuracy degradation and model re-training. We also want to highlight that our proposed AFP could effectively eliminate the computationally intensive de-quantization step existing in the dynamic quantization technique adopted by the famous machine learning frameworks (e.g., pytorch, tensorRT and etc). Moreover, we develop a framework to automatically optimize and choose the adequate AFP configuration for each layer, thus maximizing the compression efficacy. Our experiments indicate that AFP-encoded ResNet-50/MobileNet-v2 only has ~0.04/0.6% accuracy degradation w.r.t its full-precision counterpart. It outperforms the state-of-the-art works by 1.1% in accuracy using the same bit-width while reducing the energy consumption by 11.2x, which is quite impressive for inference.

Model quantization is an important mechanism for energy-efficient deployment of deep neural networks on resource-constrained devices by reducing the bit precision of weights and activations. However, it remains challenging to maintain high accuracy as bit precision decreases, especially for low-precision networks (e.g., 2-bit MobileNetV2). Existing methods have explored to address this problem by minimizing the quantization error or mimicking the data distribution of full-precision networks. In this work, we propose a novel weight regularization algorithm for improving low-precision network quantization. Instead of constraining the overall data distribution, we separably optimize all elements in each quantization bin to be as close to the target quantized value as possible. Such bin regularization (BR) mechanism encourages the weight distribution of each quantization bin to be sharp and approximate to a Dirac delta distribution ideally. Experiments demonstrate that our method achieves consistent improvements over the state-of-the-art quantization-aware training methods for different low-precision networks. Particularly, our bin regularization improves LSQ for 2-bit MobileNetV2 and MobileNetV3-Small by 3.9% and 4.9% top-1 accuracy on ImageNet, respectively.

Batch Whitening is a technique that accelerates and stabilizes training by transforming input features to have a zero mean (Centering) and a unit variance (Scaling), and by removing linear correlation between channels (Decorrelation). In commonly used structures, which are empirically optimized with Batch Normalization, the normalization layer appears between convolution and activation function. Following Batch Whitening studies have employed the same structure without further analysis; even Batch Whitening was analyzed on the premise that the input of a linear layer is whitened. To bridge the gap, we propose a new Convolutional Unit that in line with the theory, and our method generally improves the performance of Batch Whitening. Moreover, we show the inefficacy of the original Convolutional Unit by investigating rank and correlation of features. As our method is employable off-the-shelf whitening modules, we use Iterative Normalization (IterNorm), the state-of-the-art whitening module, and obtain significantly improved performance on five image classification datasets: CIFAR-10, CIFAR-100, CUB-200-2011, Stanford Dogs, and ImageNet. Notably, we verify that our method improves stability and performance of whitening when using large learning rate, group size, and iteration number.

Most existing image de-raining networks could only learn fixed mapping rules between paired rainy/clean images on single synthetic dataset and then stay static for lifetime. However, since single synthetic dataset merely provides a partial view for the distribution of rain streaks, the deep models well trained on an individual synthetic dataset tend to overfit on this biased distribution. This leads to the inability of these methods to well generalize to complex and changeable real-world rainy scenes, thus limiting their practical applications. In this paper, we try for the first time to accumulate the de-raining knowledge from multiple synthetic datasets on a single network parameter set to improve the de-raining generalization of deep networks. To achieve this goal, we explore Neural Reorganization (NR) to allow the de-raining network to keep a subtle stability-plasticity trade-off rather than naive stabilization after training phase. Specifically, we design our NR algorithm by borrowing the synaptic consolidation mechanism in the biological brain and knowledge distillation. Equipped with our NR algorithm, the deep model can be trained on a list of synthetic rainy datasets by overcoming catastrophic forgetting, making it a general-version de-raining network. Extensive experimental validation shows that due to the successful accumulation of de-raining knowledge, our proposed method can not only process multiple synthetic datasets consistently, but also achieve state-of-the-art results when dealing with real-world rainy images.

Contrastive learning, which aims at minimizing the distance between positive pairs while maximizing that of negative ones, has been widely and successfully applied in unsupervised feature learning, where the design of positive and negative (pos/neg) pairs is one of its keys. In this paper, we attempt to devise a feature-level data manipulation, differing from data augmentation, to enhance the generic contrastive self-supervised learning. To this end, we first design a visualization scheme for pos/neg score (pos/neg score indicates cosine similarity of pos/neg pair.) distribution, which enables us to analyze, interpret and understand the learning process. To our knowledge, this is the first attempt of its kind. More importantly, leveraging this tool, we gain some significant observations, which inspire our novel Feature Transformation proposals including the extrapolation of positives. This operation creates harder positives to boost the learning because hard positives enable the model to be more view-invariant. Besides, we propose the interpolation among negatives, which provides diversified negatives and makes the model more discriminative. It is the first attempt to deal with both challenges simultaneously. Experiment results show that our proposed Feature Transformation can improve at least 6.0% accuracy on ImageNet-100 over MoCo baseline, and about 2.0% accuracy on ImageNet-1K over the MoCoV2 baseline. Transferring to the downstream tasks successfully demonstrate our model is less task-bias. Visualization tools and codes: https://github.com/DTennant/CL-Visualizing-Feature-Transformation.

Though 3D object detection from point clouds has achieved rapid progress in recent years, the lack of flexible and high-performance proposal refinement remains a great hurdle for existing state-of-the-art two-stage detectors. Previous works on refining 3D proposals have relied on human-designed components such as keypoints sampling, set abstraction and multi-scale feature fusion to produce powerful 3D object representations. Such methods, however, have limited ability to capture rich contextual dependencies among points. In this paper, we leverage the high-quality region proposal network and a Channel-wise Transformer architecture to constitute our two-stage 3D object detection framework (CT3D) with minimal hand-crafted design. The proposed CT3D simultaneously performs proposal-aware embedding and channel-wise context aggregation for the point features within each proposal. Specifically, CT3D uses proposal's keypoints for spatial contextual modelling and learns attention propagation in the encoding module, mapping the proposal to point embeddings. Next, a new channel-wise decoding module enriches the query-key interaction via channel-wise re-weighting to effectively merge multi-level contexts, which contributes to more accurate object predictions. Extensive experiments demonstrate that our CT3D method has superior performance and excellent scalability. Remarkably, CT3D achieves the AP of 81.77% in the moderate car category on the KITTI test 3D detection benchmark, outperforms state-of-the-art 3D detectors.

Unsupervised pretraining has achieved great success and many recently works have shown unsupervised pretraining can achieve comparable or even slightly better transfer performance than supervised pretraining on downstream target datasets. But in this paper, we find this conclusion may not hold when the target dataset has very few labeled samples for finetuning, ie, few-label transfer. We analyze the possible reason from the clustering perspective: 1) The clustering quality of target samples is of great importance to few-label transfer; 2) Though contrastive learning is essentially to learn how to cluster, its clustering quality is still inferior to supervised pretraining due to lack of label supervision. Based on the analysis, we interestingly discover that only involving some unlabeled target domain into the unsupervised pretraining can improve the clustering quality, subsequently reducing the transfer performance gap with supervised pretraining. This finding also motivates us to propose a new progressive few-label transfer algorithm for real applications, which aims to maximize the transfer performance under a limited annotation budget. To support our analysis and proposed method, we conduct extensive experiments on nine different target datasets. Experimental results show our proposed method can significantly boost the few-label transfer performance of unsupervised pretraining.

We investigate the impact of aliasing on generalization in Deep Convolutional Networks and show that data augmentation schemes alone are unable to prevent it due to structural limitations in widely used architectures. Drawing insights from frequency analysis theory, we take a closer look at Resnet and EfficientNet architectures and review the trade-off between aliasing and information loss in each of their major components. We show how to mitigate aliasing by inserting non-trainable low-pass filters at key locations, particularly where networks lack the capacity to learn them. These simple architectural changes lead to substantial improvements in generalization on i.i.d. and even more on out-of-distribution conditions, such as image classification under natural corruptions on ImageNet-C and few-shot learning on Meta-Dataset. State-of-the art results are achieved on both datasets without introducing additional trainable parameters and using the default hyper-parameters of open source codebases.

Measuring the acoustic characteristics of a space is often done by capturing its impulse response (IR), a representation of how a full-range stimulus sound excites it. This work generates an IR from a single image, which can then be applied to other signals using convolution, simulating the reverberant characteristics of the space shown in the image. Recording these IRs is both time-intensive and expensive, and often infeasible for inaccessible locations. We use an end-to-end neural network architecture to generate plausible audio impulse responses from single images of acoustic environments. We evaluate our method both by comparisons to ground truth data and by human expert evaluation. We demonstrate our approach by generating plausible impulse responses from diverse settings and formats including well known places, musical halls, rooms in paintings, images from animations and computer games, synthetic environments generated from text, panoramic images, and video conference backgrounds.

In this paper, we present a novel approach to synthesize realistic images based on their semantic layouts. It hypothesizes that for objects with similar appearance, they share similar representation. Our method establishes dependencies between regions according to their appearance correlation, yielding both spatially variant and associated representations. Conditioning on these features, we propose a dynamic weighted network constructed by spatially conditional computation (with both convolution and normalization). More than preserving semantic distinctions, the given dynamic network strengthens semantic relevance, benefiting global structure and detail synthesis. We demonstrate that our method gives the compelling generation performance qualitatively and quantitatively with extensive experiments on benchmarks.

This paper is concerned with synthesizing images conditioned on a layout (a set of bounding boxes with object categories). Existing works construct a layout-mask-image pipeline. Object masks are generated separately and mapped to bounding boxes to form a whole semantic segmentation mask (layout-to-mask), with which a new image is generated (mask-to-image). However, overlapped boxes in layouts result in overlapped object masks, which reduces the mask clarity and causes confusion in image generation. We hypothesize the importance of generating clean and semantically clear semantic masks. The hypothesis is supported by the finding that the performance of state-of-the-art LostGAN decreases when input masks are tainted. Motivated by this hypothesis, we propose Locality-Aware Mask Adaption (LAMA) module to adapt overlapped or nearby object masks in the generation. Experimental results show our proposed model with LAMA outperforms existing approaches regarding visual fidelity and alignment with input layouts. On COCO-stuff in 256x256, our method improves the state-of-the-art FID score from 41.65 to 31.12 and the SceneFID from 22.00 to 18.64.

In this paper, we present DeepSIM, a generative model for conditional image manipulation based on a single image. We find that extensive augmentation is key for enabling single image training, and incorporate the use of thin-plate-spline (TPS) as an effective augmentation. Our network learns to map between a primitive representation of the image to the image itself. The choice of a primitive representation has an impact on the ease and expressiveness of the manipulations and can be automatic (e.g. edges), manual (e.g. segmentation) or hybrid such as edges on top of segmentations. At manipulation time, our generator allows for making complex image changes by modifying the primitive input representation and mapping it through the network. Our method is shown to achieve remarkable performance on image manipulation tasks.

We extend the task of composed image retrieval, where an input query consists of an image and short textual description of how to modify the image. Existing methods have only been applied to non-complex images within narrow domains, such as fashion products, thereby limiting the scope of study on in-depth visual reasoning in rich image and language contexts. To address this issue, we collect the Compose Image Retrieval on Real-life images (CIRR) dataset, which consists of over 36,000 pairs of crowd-sourced, open-domain images with human-generated modifying text. To extend current methods to the open-domain, we propose CIRPLANT, a transformer based model that leverages rich pre-trained vision-and-language (V&L) knowledge for modifying visual features conditioned on natural language. Retrieval is then done by nearest neighbor lookup on the modified features. We demonstrate that with a relatively simple architecture, CIRPLANT outperforms existing methods on open-domain images, while matching state-of-the-art accuracy on the existing narrow datasets, such as fashion. Together with the release of CIRR, we believe this work will inspire further research on composed image retrieval. Our dataset, code and pre-trained models are available at https://cuberick-orion.github.io/CIRR/.

The key challenge of image manipulation detection is how to learn generalizable features that are sensitive to manipulations in novel data, whilst specific to prevent false alarms on authentic images. Current research emphasizes the sensitivity, with the specificity overlooked. In this paper we address both aspects by multi-view feature learning and multi-scale supervision. By exploiting noise distribution and boundary artifact surrounding tampered regions, the former aims to learn semantic-agnostic and thus more generalizable features. The latter allows us to learn from authentic images which are nontrivial to taken into account by current semantic segmentation network based methods. Our thoughts are realized by a new network which we term MVSS-Net. Extensive experiments on five benchmark sets justify the viability of MVSS-Net for both pixel-level and image-level manipulation detection.

Deep generative approaches have recently made considerable progress in image inpainting by introducing structure priors. Due to the lack of proper interaction with image texture during structure reconstruction, however, current solutions are incompetent in handling the cases with large corruptions, and they generally suffer from distorted results. In this paper, we propose a novel two-stream network for image inpainting, which models the structure-constrained texture synthesis and texture-guided structure reconstruction in a coupled manner so that they better leverage each other for more plausible generation. Furthermore, to enhance the global consistency, a Bi-directional Gated Feature Fusion (Bi-GFF) module is designed to exchange and combine the structure and texture information and a Contextual Feature Aggregation (CFA) module is developed to refine the generated contents by region affinity learning and multi-scale feature aggregation. Qualitative and quantitative experiments on the CelebA, Paris StreetView and Places2 datasets demonstrate the superiority of the proposed method. Our code is available at https://github.com/Xiefan-Guo/CTSDG.

Image harmonization, aiming to make composite images look more realistic, is an important and challenging task. The composite, synthesized by combining foreground from one image with background from another image, inevitably suffers from the issue of inharmonious appearance caused by distinct imaging conditions, i.e., lights. Current solutions mainly adopt an encoder-decoder architecture with convolutional neural network (CNN) to capture the context of composite images, trying to understand what it looks like in the surrounding background near the foreground. In this work, we seek to solve image harmonization with Transformer, by leveraging its powerful ability of modeling long-range context dependencies, for adjusting foreground light to make it compatible with background light while keeping structure and semantics unchanged. We present the design of our harmonization Transformer frameworks without and with disentanglement, as well as comprehensive experiments and ablation study, demonstrating the power of Transformer and investigating the Transformer for vision. Our method achieves state-of-the-art performance on both image harmonization and image inpainting/enhancement, indicating its superiority. Our code and models are available at https://github.com/zhenglab/HarmonyTransformer.

Co-occurrent visual pattern makes aggregating contextual information a common paradigm to enhance the pixel representation for semantic image segmentation. The existing approaches focus on modeling the context from the perspective of the whole image, i.e., aggregating the image-level contextual information. Despite impressive, these methods weaken the significance of the pixel representations of the same category, i.e., the semantic-level contextual information. To address this, this paper proposes to augment the pixel representations by aggregating the image-level and semantic-level contextual information, respectively. First, an image-level context module is designed to capture the contextual information for each pixel in the whole image. Second, we aggregate the representations of the same category for each pixel where the category regions are learned under the supervision of the ground-truth segmentation. Third, we compute the similarities between each pixel representation and the image-level contextual information, the semantic-level contextual information, respectively. At last, a pixel representation is augmented by weighted aggregating both the image-level contextual information and the semantic-level contextual information with the similarities as the weights. Integrating the image-level and semantic-level context allows this paper to report state-of-the-art accuracy on four benchmarks, i.e., ADE20K, LIP, COCOStuff and Cityscapes.

Recently, contrastive learning has achieved great results in self-supervised learning, where the main idea is to pull two augmentations of an image (positive pairs) closer compared to other random images (negative pairs). We argue that not all negative images are equally negative. Hence, we introduce a self-supervised learning algorithm where we use a soft similarity for the negative images rather than a binary distinction between positive and negative pairs. We iteratively distill a slowly evolving teacher model to the student model by capturing the similarity of a query image to some random images and transferring that knowledge to the student. Specifically, our method should handle unbalanced and unlabeled data better than existing contrastive learning methods, because the randomly chosen negative set might include many samples that are semantically similar to the query image. In this case, our method labels them as highly similar while standard contrastive methods label them as negatives. Our method achieves comparable results to the state-of-the-art models. Our code is available here: https://github.com/UMBCvision/ISD

Denoising diffusion probabilistic models (DDPM) have shown remarkable performance in unconditional image generation. However, due to the stochasticity of the generative process in DDPM, it is challenging to generate images with the desired semantics. In this work, we propose Iterative Latent Variable Refinement (ILVR), a method to guide the generative process in DDPM to generate high-quality images based on a given reference image. Here, the refinement of the generative process in DDPM enables a single DDPM to sample images from various sets directed by the reference image. The proposed ILVR method generates high-quality images while controlling the generation. The controllability of our method allows adaptation of a single DDPM without any additional learning in various image generation tasks, such as generation from various downsampling factors, multi-domain image translation, paint-to-image, and editing with scribbles.