Search

Modern neural networks use building blocks such as convolutions that are equivariant to arbitrary 2D translations. However, these vanilla blocks are not equivariant to arbitrary 3D translations in the projective manifold. Even then, all monocular 3D detectors use vanilla blocks to obtain the 3D coordinates, a task for which the vanilla blocks are not designed for. This paper takes the first step towards convolutions equivariant to arbitrary 3D translations in the projective manifold. Since the depth is the hardest to estimate for monocular detection, this paper proposes Depth EquiVarIAnt NeTwork (DEVIANT) built with existing scale equivariant steerable blocks. As a result, DEVIANT is equivariant to the depth translations in the projective manifold whereas vanilla networks are not. The additional depth equivariance forces the DEVIANT to learn consistent depth estimates, and therefore, DEVIANT achieves state-of-the-art monocular 3D detection results on KITTI and Waymo datasets in the image-only category and performs competitively to methods using extra information. Moreover, DEVIANT works better than vanilla networks in cross-dataset evaluation. Code and models at https://github.com/abhi1kumar/DEVIANT, Comment: ECCV 2022

Recognizing scenes and objects in 3D from a single image is a longstanding goal of computer vision with applications in robotics and AR/VR. For 2D recognition, large datasets and scalable solutions have led to unprecedented advances. In 3D, existing benchmarks are small in size and approaches specialize in few object categories and specific domains, e.g. urban driving scenes. Motivated by the success of 2D recognition, we revisit the task of 3D object detection by introducing a large benchmark, called Omni3D. Omni3D re-purposes and combines existing datasets resulting in 234k images annotated with more than 3 million instances and 98 categories. 3D detection at such scale is challenging due to variations in camera intrinsics and the rich diversity of scene and object types. We propose a model, called Cube R-CNN, designed to generalize across camera and scene types with a unified approach. We show that Cube R-CNN outperforms prior works on the larger Omni3D and existing benchmarks. Finally, we prove that Omni3D is a powerful dataset for 3D object recognition and show that it improves single-dataset performance and can accelerate learning on new smaller datasets via pre-training., Comment: CVPR 2023, Project website: https://omni3d.garrickbrazil.com/

Neural network models trained on text data have been found to encode undesirable linguistic or sensitive concepts in their representation. Removing such concepts is non-trivial because of a complex relationship between the concept, text input, and the learnt representation. Recent work has proposed post-hoc and adversarial methods to remove such unwanted concepts from a model's representation. Through an extensive theoretical and empirical analysis, we show that these methods can be counter-productive: they are unable to remove the concepts entirely, and in the worst case may end up destroying all task-relevant features. The reason is the methods' reliance on a probing classifier as a proxy for the concept. Even under the most favorable conditions for learning a probing classifier when a concept's relevant features in representation space alone can provide 100% accuracy, we prove that a probing classifier is likely to use non-concept features and thus post-hoc or adversarial methods will fail to remove the concept correctly. These theoretical implications are confirmed by experiments on models trained on synthetic, Multi-NLI, and Twitter datasets. For sensitive applications of concept removal such as fairness, we recommend caution against using these methods and propose a spuriousness metric to gauge the quality of the final classifier.

Non-orthogonal multiple access (NOMA) is recognized as a promising radio access technique for the next generation wireless systems. We consider a practical downlink NOMA system with imperfect successive interference cancellation and derive bounds on the power allocation factors for a given number of users in each cluster. We propose a minimum signal-to-interference-plus-noise ratio difference criterion between two successive NOMA users in a cluster of users to achieve higher rates than an equivalent orthogonal multiple access (OMA) system. We then propose adaptive multi-user clustering and power allocation algorithms for downlink NOMA systems. Through extensive simulations, we show that the proposed algorithms achieve higher rates than the state-of-the-art algorithms.

Rate-Splitting multiple access (RSMA) has emerged as a key enabler in improving the performance of the beyond fifth-generation (5G) cellular networks. The existing literature has typically considered the sum rate of the users to evaluate the performance of RSMA. However, it has been shown in the existing works that maximizing the sum rate can result in asymmetric user performance. It significantly enhances one user's rate at the cost of the rate of another RSMA user. Further, imperfections can reduce the performance of successive interference cancellation (SIC)-based RSMA. Therefore, in this letter, we consider the imperfection in SIC and derive suitable bounds on fractions of the power allocated for common and private messages and the fraction of common message intended for each user in an RSMA pair such that their individual RSMA rates are greater than their respective orthogonal multiple access (OMA) rates. Through simulations, we validate the derived bounds. We show that they can be used to appropriately select the RSMA parameters resulting in users' RSMA rates being better than their respective OMA rates., Comment: To be submitted to IEEE Communications Letters

Non-orthogonal multiple access (NOMA) is considered as one of the predominant multiple access technique for the next-generation cellular networks. We consider a 2-user pair downlink NOMA system with imperfect successive interference cancellation (SIC). We consider bounds on the power allocation factors and then formulate the power allocation as an optimization problem to achieve {$\alpha$-Fairness} among the paired users. We show that {$\alpha$-Fairness} based power allocation factor coincides with lower bound on power allocation factor in case of perfect SIC and $\alpha > 2$. Further, as long as the proposed criterion is satisfied, it converges to the upper bound with increasing imperfection in SIC. Similarly, we show that, for $0<\alpha<1$, the optimal power allocation factor coincides with the derived lower bound on power allocation. Based on these observations, we then propose a low complexity sub-optimal algorithm. Through extensive simulations, we analyse the performance of the proposed algorithm and compare the performance against the state-of-the-art algorithms. We show that even though Near-Far based pairing achieves better fairness than the proposed algorithms, it fails to achieve rates equivalent to its orthogonal multiple access counterparts with increasing imperfections in SIC. Further, we show that the proposed optimal and sub-optimal algorithms achieve significant improvements in terms of fairness as compared to the state-of-the-art algorithms.

Non-orthogonal multiple access (NOMA) is a promising multiple access technology to improve the throughput and spectral efficiency of the users for 5G and beyond cellular networks. Similarly, coordinated multi-point transmission and reception (CoMP) is an existing technology to improve the coverage of cell-edge users. Hence, NOMA along with CoMP can potentially enhance the throughput and coverage of the users. However, the order of implementation of CoMP and NOMA can have a significant impact on the system performance of Ultra-dense networks (UDNs). Motivated by this, we study the performance of the CoMP and NOMA based UDN by proposing two kinds of user grouping and pairing schemes that differ in the order in which CoMP and NOMA are performed for a group of users. Detailed simulation results are presented comparing the proposed schemes with the state-of-the-art systems with varying user and base station densities. Through numerical results, we show that the proposed schemes can be used to achieve a suitable coverage-throughout trade-off in UDNs., Comment: Submitted to WCNC Conference workshops

Non-orthogonal multiple access (NOMA) is considered a key technology for improving the spectral efficiency of fifth-generation (5G) and beyond 5G cellular networks. NOMA is beneficial when the channel vectors of the users are in the same direction, which is not always possible in conventional wireless systems. With the help of a reconfigurable intelligent surface (RIS), the base station can control the directions of the channel vectors of the users. Thus, by combining both technologies, the RIS-assisted NOMA systems are expected to achieve greater improvements in the network throughput. However, ideal phase control at the RIS is unrealizable in practice because of the imperfections in the channel estimations and the hardware limitations. This imperfection in phase control can have a significant impact on the system performance. Motivated by this, in this paper, we consider an RIS-assisted uplink NOMA system in the presence of imperfect phase compensation. We formulate the criterion for pairing the users that achieves minimum required data rates. We propose adaptive user pairing algorithms that maximize spectral or energy efficiency. We then derive various bounds on power allocation factors for the paired users. Through extensive simulation results, we show that the proposed algorithms significantly outperform the state-of-the-art algorithms in terms of spectral and energy efficiency.