Shizhong Han

Minghua Liu, Yinhao Zhu, Hong Cai et al.

Generalizable 3D part segmentation is important but challenging in vision and robotics. Training deep models via conventional supervised methods requires large-scale 3D datasets with fine-grained part annotations, which are costly to collect. This paper explores an alternative way for low-shot part segmentation of 3D point clouds by leveraging a pretrained image-language model, GLIP, which achieves superior performance on open-vocabulary 2D detection. We transfer the rich knowledge from 2D to 3D through GLIP-based part detection on point cloud rendering and a novel 2D-to-3D label lifting algorithm. We also utilize multi-view 3D priors and few-shot prompt tuning to boost performance significantly. Extensive evaluation on PartNet and PartNet-Mobility datasets shows that our method enables excellent zero-shot 3D part segmentation. Our few-shot version not only outperforms existing few-shot approaches by a large margin but also achieves highly competitive results compared to the fully supervised counterpart. Furthermore, we demonstrate that our method can be directly applied to iPhone-scanned point clouds without significant domain gaps.

Minghan Zhu, Shizhong Han, Hong Cai et al.

In this paper, we develop rotation-equivariant neural networks for 4D panoptic segmentation. 4D panoptic segmentation is a benchmark task for autonomous driving that requires recognizing semantic classes and object instances on the road based on LiDAR scans, as well as assigning temporally consistent IDs to instances across time. We observe that the driving scenario is symmetric to rotations on the ground plane. Therefore, rotation-equivariance could provide better generalization and more robust feature learning. Specifically, we review the object instance clustering strategies and restate the centerness-based approach and the offset-based approach as the prediction of invariant scalar fields and equivariant vector fields. Other sub-tasks are also unified from this perspective, and different invariant and equivariant layers are designed to facilitate their predictions. Through evaluation on the standard 4D panoptic segmentation benchmark of SemanticKITTI, we show that our equivariant models achieve higher accuracy with lower computational costs compared to their non-equivariant counterparts. Moreover, our method sets the new state-of-the-art performance and achieves 1st place on the SemanticKITTI 4D Panoptic Segmentation leaderboard.

Pierre-David Letourneau, Manish Kumar Singh, Hsin-Pai Cheng et al.

We present Polynomial Attention Drop-in Replacement (PADRe), a novel and unifying framework designed to replace the conventional self-attention mechanism in transformer models. Notably, several recent alternative attention mechanisms, including Hyena, Mamba, SimA, Conv2Former, and Castling-ViT, can be viewed as specific instances of our PADRe framework. PADRe leverages polynomial functions and draws upon established results from approximation theory, enhancing computational efficiency without compromising accuracy. PADRe's key components include multiplicative nonlinearities, which we implement using straightforward, hardware-friendly operations such as Hadamard products, incurring only linear computational and memory costs. PADRe further avoids the need for using complex functions such as Softmax, yet it maintains comparable or superior accuracy compared to traditional self-attention. We assess the effectiveness of PADRe as a drop-in replacement for self-attention across diverse computer vision tasks. These tasks include image classification, image-based 2D object detection, and 3D point cloud object detection. Empirical results demonstrate that PADRe runs significantly faster than the conventional self-attention (11x ~ 43x faster on server GPU and mobile NPU) while maintaining similar accuracy when substituting self-attention in the transformer models.

Rajeev Yasarla, Deepti Hegde, Shizhong Han et al.

Vision-Language-Action (VLA) models are emerging as highly effective planning models for end-to-end autonomous driving systems. However, current works mostly rely on imitation learning from sparse trajectory annotations and under-utilize their potential as generative models. We propose Generative Scenario Rollouts (GeRo), a plug-and-play framework for VLA models that jointly performs planning and generation of language-grounded future traffic scenes through an autoregressive rollout strategy. First, a VLA model is trained to encode ego vehicle and agent dynamics into latent tokens under supervision from planning, motion, and language tasks, facilitating text-aligned generation. Next, GeRo performs language-conditioned autoregressive generation. Given multi-view images, a scenario description, and ego-action questions, it generates future latent tokens and textual responses to guide long-horizon rollouts. A rollout-consistency loss stabilizes predictions using ground truth or pseudo-labels, mitigating drift and preserving text-action alignment. This design enables GeRo to perform temporally consistent, language-grounded rollouts that support long-horizon reasoning and multi-agent planning. On Bench2Drive, GeRo improves driving score and success rate by +15.7 and +26.2, respectively. By integrating reinforcement learning with generative rollouts, GeRo achieves state-of-the-art closed-loop and open-loop performance, demonstrating strong zero-shot robustness. These results highlight the promise of generative, language-conditioned reasoning as a foundation for safer and more interpretable end-to-end autonomous driving.

Rajeev Yasarla, Deepti Hegde, Hsin-Pai Cheng et al.

Vision-language-action (VLA) models are effective as end-to-end motion planners, but can be brittle when evaluated in closed-loop settings due to being trained under traditional imitation learning framework. Existing closed-loop supervision approaches lack scalability and fail to completely model a reactive environment. We propose MAPLE, a novel framework for reactive, multi-agent rollout of a dynamic driving scenario in the latent space of the VLA model. The ego vehicle and nearby traffic agents are independently controlled over multi-step horizons, while being reactive to other agents in the scene, enabling closed-loop training. MAPLE consists of two training stages: (1) supervised fine-tuning on the latent rollouts based on ground-truth trajectories, followed by (2) reinforcement learning with global and agent -specific rewards that encourage safety, progress, and interaction realism. We further propose diversity rewards that encourage the model to generate planning behaviors that may not be present in logged driving data. Notably, our closed-loop training framework is scalable and does not require external simulators, which can be computationally expensive to run and have limited visual fidelity to the real-world. MAPLE achieves state-of-the-art driving performance on Bench2Drive and demonstrates scalable, closed-loop multi-agent play for robust E2E autonomous driving systems.

Zhuojin Li, Hsin-Pai Cheng, Hong Cai et al.

Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retained tokens. We further introduce a progressive, block-adaptive pruning schedule that increases pruning gradually and allocates larger budgets to blocks and denoising steps with higher redundancy. Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality. Notably, CoReDiT also increases on-device memory head-room, enabling higher-resolution generation.

Shiva Afshar, Yinghan Chen, Shizhong Han et al.

Combining multiple predictors obtained from distributed data sources to an accurate meta-learner is promising to achieve enhanced performance in lots of prediction problems. As the accuracy of each predictor is usually unknown, integrating the predictors to achieve better performance is challenging. Conventional ensemble learning methods assess the accuracy of predictors based on extensive labeled data. In practical applications, however, the acquisition of such labeled data can prove to be an arduous task. Furthermore, the predictors under consideration may exhibit high degrees of correlation, particularly when similar data sources or machine learning algorithms were employed during their model training. In response to these challenges, this paper introduces a novel structured unsupervised ensemble learning model (SUEL) to exploit the dependency between a set of predictors with continuous predictive scores, rank the predictors without labeled data and combine them to an ensembled score with weights. Two novel correlation-based decomposition algorithms are further proposed to estimate the SUEL model, constrained quadratic optimization (SUEL.CQO) and matrix-factorization-based (SUEL.MF) approaches. The efficacy of the proposed methods is rigorously assessed through both simulation studies and real-world application of risk genes discovery. The results compellingly demonstrate that the proposed methods can efficiently integrate the dependent predictors to an ensemble model without the need of ground truth data.

Deepti Hegde, Rajeev Yasarla, Hong Cai et al.

Autonomous driving demands safe motion planning, especially in critical "long-tail" scenarios. Recent end-to-end autonomous driving systems leverage large language models (LLMs) as planners to improve generalizability to rare events. However, using LLMs at test time introduces high computational costs. To address this, we propose DiMA, an end-to-end autonomous driving system that maintains the efficiency of an LLM-free (or vision-based) planner while leveraging the world knowledge of an LLM. DiMA distills the information from a multi-modal LLM to a vision-based end-to-end planner through a set of specially designed surrogate tasks. Under a joint training strategy, a scene encoder common to both networks produces structured representations that are semantically grounded as well as aligned to the final planning objective. Notably, the LLM is optional at inference, enabling robust planning without compromising on efficiency. Training with DiMA results in a 37% reduction in the L2 trajectory error and an 80% reduction in the collision rate of the vision-based planner, as well as a 44% trajectory error reduction in longtail scenarios. DiMA also achieves state-of-the-art performance on the nuScenes planning benchmark.

Rajeev Yasarla, Shizhong Han, Hsin-Pai Cheng et al.

End-to-end (E2E) autonomous driving has recently emerged as a new paradigm, offering significant potential. However, few studies have looked into the practical challenge of deployment across domains (e.g., cities). Although several works have incorporated Large Language Models (LLMs) to leverage their open-world knowledge, LLMs do not guarantee cross-domain driving performance and may incur prohibitive retraining costs during domain adaptation. In this paper, we propose RoCA, a novel framework for robust cross-domain E2E autonomous driving. RoCA formulates the joint probabilistic distribution over the tokens that encode ego and surrounding vehicle information in the E2E pipeline. Instantiating with a Gaussian process (GP), RoCA learns a set of basis tokens with corresponding trajectories, which span diverse driving scenarios. Then, given any driving scene, it is able to probabilistically infer the future trajectory. By using RoCA together with a base E2E model in source-domain training, we improve the generalizability of the base model, without requiring extra inference computation. In addition, RoCA enables robust adaptation on new target domains, significantly outperforming direct finetuning. We extensively evaluate RoCA on various cross-domain scenarios and show that it achieves strong domain generalization and adaptation performance.

Yunxiao Shi, Yinhao Zhu, Shizhong Han et al.

Occupancy prediction infers fine-grained 3D geometry and semantics from camera images of the surrounding environment, making it a critical perception task for autonomous driving. Existing methods either adopt dense grids as scene representation, which is difficult to scale to high resolution, or learn the entire scene using a single set of sparse queries, which is insufficient to handle the various object characteristics. In this paper, we present ODG, a hierarchical dual sparse Gaussian representation to effectively capture complex scene dynamics. Building upon the observation that driving scenes can be universally decomposed into static and dynamic counterparts, we define dual Gaussian queries to better model the diverse scene objects. We utilize a hierarchical Gaussian transformer to predict the occupied voxel centers and semantic classes along with the Gaussian parameters. Leveraging the real-time rendering capability of 3D Gaussian Splatting, we also impose rendering supervision with available depth and semantic map annotations injecting pixel-level alignment to boost occupancy learning. Extensive experiments on the Occ3D-nuScenes and Occ3D-Waymo benchmarks demonstrate our proposed method sets new state-of-the-art results while maintaining low inference cost.

Rajeev Yasarla, Shizhong Han, Hong Cai et al.

Camera-based 3D object detection in Bird's Eye View (BEV) is one of the most important perception tasks in autonomous driving. Earlier methods rely on dense BEV features, which are costly to construct. More recent works explore sparse query-based detection. However, they still require a large number of queries and can become expensive to run when more video frames are used. In this paper, we propose DySS, a novel method that employs state-space learning and dynamic queries. More specifically, DySS leverages a state-space model (SSM) to sequentially process the sampled features over time steps. In order to encourage the model to better capture the underlying motion and correspondence information, we introduce auxiliary tasks of future prediction and masked reconstruction to better train the SSM. The state of the SSM then provides an informative yet efficient summarization of the scene. Based on the state-space learned features, we dynamically update the queries via merge, remove, and split operations, which help maintain a useful, lean set of detection queries throughout the network. Our proposed DySS achieves both superior detection performance and efficient inference. Specifically, on the nuScenes test split, DySS achieves 65.31 NDS and 57.4 mAP, outperforming the latest state of the art. On the val split, DySS achieves 56.2 NDS and 46.2 mAP, as well as a real-time inference speed of 33 FPS.

Yunxiao Shi, Hong Cai, Jisoo Jeong et al.

3D occupancy provides fine-grained 3D geometry and semantics for scene understanding which is critical for autonomous driving. Most existing methods, however, carry high compute costs, requiring dense 3D feature volume and cross-attention to effectively aggregate information. More recent works have adopted Bird's Eye View (BEV) or sparse points as scene representation with much reduced cost, but still suffer from their respective shortcomings. More concretely, BEV struggles with small objects that often experience significant information loss after being projected to the ground plane. On the other hand, points can flexibly model little objects in 3D, but is inefficient at capturing flat surfaces or large objects. To address these challenges, in this paper, we present a novel 3D occupancy prediction approach, BePo, which combines BEV and sparse points based representations. We propose a dual-branch design: a query-based sparse points branch and a BEV branch. The 3D information learned in the sparse points branch is shared with the BEV stream via cross-attention, which enriches the weakened signals of difficult objects on the BEV plane. The outputs of both branches are finally fused to generate predicted 3D occupancy. We conduct extensive experiments on the Occ3D-nuScenes and Occ3D-Waymo benchmarks that demonstrate the superiority of our proposed BePo. Moreover, BePo also delivers competitive inference speed when compared to the latest efficient approaches.

Shizhong Han, Hsin-Pai Cheng, Hong Cai et al.

Existing LiDAR 3D object detection methods predominantely rely on sparse convolutions and/or transformers, which can be challenging to run on resource-constrained edge devices, due to irregular memory access patterns and high computational costs. In this paper, we propose FALO, a hardware-friendly approach to LiDAR 3D detection, which offers both state-of-the-art (SOTA) detection accuracy and fast inference speed. More specifically, given the 3D point cloud and after voxelization, FALO first arranges sparse 3D voxels into a 1D sequence based on their coordinates and proximity. The sequence is then processed by our proposed ConvDotMix blocks, consisting of large-kernel convolutions, Hadamard products, and linear layers. ConvDotMix provides sufficient mixing capability in both spatial and embedding dimensions, and introduces higher-order nonlinear interaction among spatial features. Furthermore, when going through the ConvDotMix layers, we introduce implicit grouping, which balances the tensor dimensions for more efficient inference and takes into account the growing receptive field. All these operations are friendly to run on resource-constrained platforms and proposed FALO can readily deploy on compact, embedded devices. Our extensive evaluation on LiDAR 3D detection benchmarks such as nuScenes and Waymo shows that FALO achieves competitive performance. Meanwhile, FALO is 1.6~9.8x faster than the latest SOTA on mobile Graphics Processing Unit (GPU) and mobile Neural Processing Unit (NPU).

Rajeev Yasarla, Manish Kumar Singh, Hong Cai et al.

In this paper, we propose a novel video depth estimation approach, FutureDepth, which enables the model to implicitly leverage multi-frame and motion cues to improve depth estimation by making it learn to predict the future at training. More specifically, we propose a future prediction network, F-Net, which takes the features of multiple consecutive frames and is trained to predict multi-frame features one time step ahead iteratively. In this way, F-Net learns the underlying motion and correspondence information, and we incorporate its features into the depth decoding process. Additionally, to enrich the learning of multiframe correspondence cues, we further leverage a reconstruction network, R-Net, which is trained via adaptively masked auto-encoding of multiframe feature volumes. At inference time, both F-Net and R-Net are used to produce queries to work with the depth decoder, as well as a final refinement network. Through extensive experiments on several benchmarks, i.e., NYUDv2, KITTI, DDAD, and Sintel, which cover indoor, driving, and open-domain scenarios, we show that FutureDepth significantly improves upon baseline models, outperforms existing video depth estimation methods, and sets new state-of-the-art (SOTA) accuracy. Furthermore, FutureDepth is more efficient than existing SOTA video depth estimation models and has similar latencies when comparing to monocular models

Minghua Liu, Ruoxi Shi, Kaiming Kuang et al.

We introduce OpenShape, a method for learning multi-modal joint representations of text, image, and point clouds. We adopt the commonly used multi-modal contrastive learning framework for representation alignment, but with a specific focus on scaling up 3D representations to enable open-world 3D shape understanding. To achieve this, we scale up training data by ensembling multiple 3D datasets and propose several strategies to automatically filter and enrich noisy text descriptions. We also explore and compare strategies for scaling 3D backbone networks and introduce a novel hard negative mining module for more efficient training. We evaluate OpenShape on zero-shot 3D classification benchmarks and demonstrate its superior capabilities for open-world recognition. Specifically, OpenShape achieves a zero-shot accuracy of 46.8% on the 1,156-category Objaverse-LVIS benchmark, compared to less than 10% for existing methods. OpenShape also achieves an accuracy of 85.3% on ModelNet40, outperforming previous zero-shot baseline methods by 20% and performing on par with some fully-supervised methods. Furthermore, we show that our learned embeddings encode a wide range of visual and semantic concepts (e.g., subcategories, color, shape, style) and facilitate fine-grained text-3D and image-3D interactions. Due to their alignment with CLIP embeddings, our learned shape representations can also be integrated with off-the-shelf CLIP-based models for various applications, such as point cloud captioning and point cloud-conditioned image generation.

Xudong Wang, Shizhong Han, Yunqiang Chen et al.

A volumetric attention(VA) module for 3D medical image segmentation and detection is proposed. VA attention is inspired by recent advances in video processing, enables 2.5D networks to leverage context information along the z direction, and allows the use of pretrained 2D detection models when training data is limited, as is often the case for medical applications. Its integration in the Mask R-CNN is shown to enable state-of-the-art performance on the Liver Tumor Segmentation (LiTS) Challenge, outperforming the previous challenge winner by 3.9 points and achieving top performance on the LiTS leader board at the time of paper submission. Detection experiments on the DeepLesion dataset also show that the addition of VA to existing object detectors enables a 69.1 sensitivity at 0.5 false positive per image, outperforming the best published results by 6.6 points.

Jacob C. Reinhold, Yufan He, Shizhong Han et al.

Medical images are increasingly used as input to deep neural networks to produce quantitative values that aid researchers and clinicians. However, standard deep neural networks do not provide a reliable measure of uncertainty in those quantitative values. Recent work has shown that using dropout during training and testing can provide estimates of uncertainty. In this work, we investigate using dropout to estimate epistemic and aleatoric uncertainty in a CT-to-MR image translation task. We show that both types of uncertainty are captured, as defined, providing confidence in the output uncertainty estimates.

Jacob C. Reinhold, Yufan He, Shizhong Han et al.

Medical images are often used to detect and characterize pathology and disease; however, automatically identifying and segmenting pathology in medical images is challenging because the appearance of pathology across diseases varies widely. To address this challenge, we propose a Bayesian deep learning method that learns to translate healthy computed tomography images to magnetic resonance images and simultaneously calculates voxel-wise uncertainty. Since high uncertainty occurs in pathological regions of the image, this uncertainty can be used for unsupervised anomaly segmentation. We show encouraging experimental results on an unsupervised anomaly segmentation task by combining two types of uncertainty into a novel quantity we call scibilic uncertainty.

Yuchen Xu, Olivia Tang, Yucheng Tang et al.

Segmentation of abdominal computed tomography(CT) provides spatial context, morphological properties, and a framework for tissue-specific radiomics to guide quantitative Radiological assessment. A 2015 MICCAI challenge spurred substantial innovation in multi-organ abdominal CT segmentation with both traditional and deep learning methods. Recent innovations in deep methods have driven performance toward levels for which clinical translation is appealing. However, continued cross-validation on open datasets presents the risk of indirect knowledge contamination and could result in circular reasoning. Moreover, 'real world' segmentations can be challenging due to the wide variability of abdomen physiology within patients. Herein, we perform two data retrievals to capture clinically acquired deidentified abdominal CT cohorts with respect to a recently published variation on 3D U-Net (baseline algorithm). First, we retrieved 2004 deidentified studies on 476 patients with diagnosis codes involving spleen abnormalities (cohort A). Second, we retrieved 4313 deidentified studies on 1754 patients without diagnosis codes involving spleen abnormalities (cohort B). We perform prospective evaluation of the existing algorithm on both cohorts, yielding 13% and 8% failure rate, respectively. Then, we identified 51 subjects in cohort A with segmentation failures and manually corrected the liver and gallbladder labels. We re-trained the model adding the manual labels, resulting in performance improvement of 9% and 6% failure rate for the A and B cohorts, respectively. In summary, the performance of the baseline on the prospective cohorts was similar to that on previously published datasets. Moreover, adding data from the first cohort substantively improved performance when evaluated on the second withheld validation cohort.

Yuchen Xu, Olivia Tang, Yucheng Tang et al.

Abdominal multi-organ segmentation of computed tomography (CT) images has been the subject of extensive research interest. It presents a substantial challenge in medical image processing, as the shape and distribution of abdominal organs can vary greatly among the population and within an individual over time. While continuous integration of novel datasets into the training set provides potential for better segmentation performance, collection of data at scale is not only costly, but also impractical in some contexts. Moreover, it remains unclear what marginal value additional data have to offer. Herein, we propose a single-pass active learning method through human quality assurance (QA). We built on a pre-trained 3D U-Net model for abdominal multi-organ segmentation and augmented the dataset either with outlier data (e.g., exemplars for which the baseline algorithm failed) or inliers (e.g., exemplars for which the baseline algorithm worked). The new models were trained using the augmented datasets with 5-fold cross-validation (for outlier data) and withheld outlier samples (for inlier data). Manual labeling of outliers increased Dice scores with outliers by 0.130, compared to an increase of 0.067 with inliers (p<0.001, two-tailed paired t-test). By adding 5 to 37 inliers or outliers to training, we find that the marginal value of adding outliers is higher than that of adding inliers. In summary, improvement on single-organ performance was obtained without diminishing multi-organ performance or significantly increasing training time. Hence, identification and correction of baseline failure cases present an effective and efficient method of selecting training data to improve algorithm performance.

Yuankai Huo, Yucheng Tang, Yunqiang Chen et al.

Tissue window filtering has been widely used in deep learning for computed tomography (CT) image analyses to improve training performance (e.g., soft tissue windows for abdominal CT). However, the effectiveness of tissue window normalization is questionable since the generalizability of the trained model might be further harmed, especially when such models are applied to new cohorts with different CT reconstruction kernels, contrast mechanisms, dynamic variations in the acquisition, and physiological changes. We evaluate the effectiveness of both with and without using soft tissue window normalization on multisite CT cohorts. Moreover, we propose a stochastic tissue window normalization (SWN) method to improve the generalizability of tissue window normalization. Different from the random sampling, the SWN method centers the randomization around the soft tissue window to maintain the specificity for abdominal organs. To evaluate the performance of different strategies, 80 training and 453 validation and testing scans from six datasets are employed to perform multi-organ segmentation using standard 2D U-Net. The six datasets cover the scenarios, where the training and testing scans are from (1) same scanner and same population, (2) same CT contrast but different pathology, and (3) different CT contrast and pathology. The traditional soft tissue window and nonwindowed approaches achieved better performance on (1). The proposed SWN achieved general superior performance on (2) and (3) with statistical analyses, which offers better generalizability for a trained model.

Yucheng Tang, Ho Hin Lee, Yuchen Xu et al.

Dynamic contrast enhanced computed tomography (CT) is an imaging technique that provides critical information on the relationship of vascular structure and dynamics in the context of underlying anatomy. A key challenge for image processing with contrast enhanced CT is that phase discrepancies are latent in different tissues due to contrast protocols, vascular dynamics, and metabolism variance. Previous studies with deep learning frameworks have been proposed for classifying contrast enhancement with networks inspired by computer vision. Here, we revisit the challenge in the context of whole abdomen contrast enhanced CTs. To capture and compensate for the complex contrast changes, we propose a novel discriminator in the form of a multi-domain disentangled representation learning network. The goal of this network is to learn an intermediate representation that separates contrast enhancement from anatomy and enables classification of images with varying contrast time. Briefly, our unpaired contrast disentangling GAN(CD-GAN) Discriminator follows the ResNet architecture to classify a CT scan from different enhancement phases. To evaluate the approach, we trained the enhancement phase classifier on 21060 slices from two clinical cohorts of 230 subjects. Testing was performed on 9100 slices from 30 independent subjects who had been imaged with CT scans from all contrast phases. Performance was quantified in terms of the multi-class normalized confusion matrix. The proposed network significantly improved correspondence over baseline UNet, ResNet50 and StarGAN performance of accuracy scores 0.54. 0.55, 0.62 and 0.91, respectively. The proposed discriminator from the disentangled network presents a promising technique that may allow deeper modeling of dynamic imaging against patient specific anatomies.

Ho Hin Lee, Yucheng Tang, Olivia Tang et al.

Human in-the-loop quality assurance (QA) is typically performed after medical image segmentation to ensure that the systems are performing as intended, as well as identifying and excluding outliers. By performing QA on large-scale, previously unlabeled testing data, categorical QA scores can be generatedIn this paper, we propose a semi-supervised multi-organ segmentation deep neural network consisting of a traditional segmentation model generator and a QA involved discriminator. A large-scale dataset of 2027 volumes are used to train the generator, whose 2-D montage images and segmentation mask with QA scores are used to train the discriminator. To generate the QA scores, the 2-D montage images were reviewed manually and coded 0 (success), 1 (errors consistent with published performance), and 2 (gross failure). Then, the ResNet-18 network was trained with 1623 montage images in equal distribution of all three code labels and achieved an accuracy 94% for classification predictions with 404 montage images withheld for the test cohort. To assess the performance of using the QA supervision, the discriminator was used as a loss function in a multi-organ segmentation pipeline. The inclusion of QA-loss function boosted performance on the unlabeled test dataset from 714 patients to 951 patients over the baseline model. Additionally, the number of failures decreased from 606 (29.90%) to 402 (19.83%). The contributions of the proposed method are threefold: We show that (1) the QA scores can be used as a loss function to perform semi-supervised learning for unlabeled data, (2) the well trained discriminator is learnt by QA score rather than traditional true/false, and (3) the performance of multi-organ segmentation on unlabeled datasets can be fine-tuned with more robust and higher accuracy than the original baseline method.

Jie Cai, Zibo Meng, Ahmed Shehab Khan et al.

Emotion recognition plays an important role in human-computer interaction (HCI) and has been extensively studied for decades. Although tremendous improvements have been achieved for posed expressions, recognizing human emotions in "close-to-real-world" environments remains a challenge. In this paper, we proposed two strategies to fuse information extracted from different modalities, i.e., audio and visual. Specifically, we utilized LBP-TOP, an ensemble of CNNs, and a bi-directional LSTM (BLSTM) to extract features from the visual channel, and the OpenSmile toolkit to extract features from the audio channel. Two kinds of fusion methods, i,e., feature-level fusion and model-level fusion, were developed to utilize the information extracted from the two channels. Experimental results on the EmotiW2018 AFEW dataset have shown that the proposed fusion methods outperform the baseline methods significantly and achieve better or at least comparable performance compared with the state-of-the-art methods, where the model-level fusion performs better when one of the channels totally fails.

Jie Cai, Zibo Meng, Ahmed Shehab Khan et al.

A novel Identity-Free conditional Generative Adversarial Network (IF-GAN) was proposed for Facial Expression Recognition (FER) to explicitly reduce high inter-subject variations caused by identity-related facial attributes, e.g., age, race, and gender. As part of an end-to-end system, a cGAN was designed to transform a given input facial expression image to an "average" identity face with the same expression as the input. Then, identity-free FER is possible since the generated images have the same synthetic "average" identity and differ only in their displayed expressions. Experiments on four facial expression datasets, one with spontaneous expressions, show that IF-GAN outperforms the baseline CNN and achieves state-of-the-art performance for FER.

Shizhong Han, Zibo Meng, Zhiyuan Li et al.

Recognizing facial action units (AUs) during spontaneous facial displays is a challenging problem. Most recently, Convolutional Neural Networks (CNNs) have shown promise for facial AU recognition, where predefined and fixed convolution filter sizes are employed. In order to achieve the best performance, the optimal filter size is often empirically found by conducting extensive experimental validation. Such a training process suffers from expensive training cost, especially as the network becomes deeper. This paper proposes a novel Optimized Filter Size CNN (OFS-CNN), where the filter sizes and weights of all convolutional layers are learned simultaneously from the training data along with learning convolution filters. Specifically, the filter size is defined as a continuous variable, which is optimized by minimizing the training loss. Experimental results on two AU-coded spontaneous databases have shown that the proposed OFS-CNN is capable of estimating optimal filter size for varying image resolution and outperforms traditional CNNs with the best filter size obtained by exhaustive search. The OFS-CNN also beats the CNN using multiple filter sizes and more importantly, is much more efficient during testing with the proposed forward-backward propagation algorithm.

Shizhong Han, Zibo Meng, Ahmed Shehab Khan et al.

Recognizing facial action units (AUs) from spontaneous facial expressions is still a challenging problem. Most recently, CNNs have shown promise on facial AU recognition. However, the learned CNNs are often overfitted and do not generalize well to unseen subjects due to limited AU-coded training images. We proposed a novel Incremental Boosting CNN (IB-CNN) to integrate boosting into the CNN via an incremental boosting layer that selects discriminative neurons from the lower layer and is incrementally updated on successive mini-batches. In addition, a novel loss function that accounts for errors from both the incremental boosted classifier and individual weak classifiers was proposed to fine-tune the IB-CNN. Experimental results on four benchmark AU databases have demonstrated that the IB-CNN yields significant improvement over the traditional CNN and the boosting CNN without incremental learning, as well as outperforming the state-of-the-art CNN-based methods in AU recognition. The improvement is more impressive for the AUs that have the lowest frequencies in the databases.

Zibo Meng, Shizhong Han, Ping Liu et al.

It is challenging to recognize facial action unit (AU) from spontaneous facial displays, especially when they are accompanied by speech. The major reason is that the information is extracted from a single source, i.e., the visual channel, in the current practice. However, facial activity is highly correlated with voice in natural human communications. Instead of solely improving visual observations, this paper presents a novel audiovisual fusion framework, which makes the best use of visual and acoustic cues in recognizing speech-related facial AUs. In particular, a dynamic Bayesian network (DBN) is employed to explicitly model the semantic and dynamic physiological relationships between AUs and phonemes as well as measurement uncertainty. A pilot audiovisual AU-coded database has been collected to evaluate the proposed framework, which consists of a "clean" subset containing frontal faces under well controlled circumstances and a challenging subset with large head movements and occlusions. Experiments on this database have demonstrated that the proposed framework yields significant improvement in recognizing speech-related AUs compared to the state-of-the-art visual-based methods especially for those AUs whose visual observations are impaired during speech, and more importantly also outperforms feature-level fusion methods by explicitly modeling and exploiting physiological relationships between AUs and phonemes.

Zibo Meng, Shizhong Han, Yan Tong

Extensive efforts have been devoted to recognizing facial action units (AUs). However, it is still challenging to recognize AUs from spontaneous facial displays especially when they are accompanied with speech. Different from all prior work that utilized visual observations for facial AU recognition, this paper presents a novel approach that recognizes speech-related AUs exclusively from audio signals based on the fact that facial activities are highly correlated with voice during speech. Specifically, dynamic and physiological relationships between AUs and phonemes are modeled through a continuous time Bayesian network (CTBN); then AU recognition is performed by probabilistic inference via the CTBN model. A pilot audiovisual AU-coded database has been constructed to evaluate the proposed audio-based AU recognition framework. The database consists of a "clean" subset with frontal and neutral faces and a challenging subset collected with large head movements and occlusions. Experimental results on this database show that the proposed CTBN model achieves promising recognition performance for 7 speech-related AUs and outperforms the state-of-the-art visual-based methods especially for those AUs that are activated at low intensities or "hardly visible" in the visual channel. Furthermore, the CTBN model yields more impressive recognition performance on the challenging subset, where the visual-based approaches suffer significantly.