Previous cycle-consistency correspondence learning methods usually leverage image patches for training. In this paper, we present a fully convolutional method, which is simpler and more coherent to the inference process. While directly applying fully
convolutional training results in model collapse, we study the underline reason behind this collapse phenomenon, indicating that the absolute positions of pixels provide a shortcut to easily accomplish cycle-consistence, which hinders the learning of meaningful visual representations. To break this absolute position shortcut, we propose to apply different crops for forward and backward frames, and adopt feature warping to establish correspondence between two crops of a same frame. The former technique enforces the corresponding pixels at forward and back tracks to have different absolute positions, and the latter effectively blocks the shortcuts going between forward and back tracks. In three label propagation benchmarks for pose tracking, face landmark tracking and video object segmentation, our method largely improves the results of vanilla fully convolutional cycle-consistency method, achieving very competitive performance compared with the self-supervised state-of-the-art approaches.
We are witnessing a modeling shift from CNN to Transformers in computer vision. In this work, we present a self-supervised learning approach called MoBY, with Vision Transformers as its backbone architecture. The approach basically has no new inventi
ons, which is combined from MoCo v2 and BYOL and tuned to achieve reasonably high accuracy on ImageNet-1K linear evaluation: 72.8% and 75.0% top-1 accuracy using DeiT-S and Swin-T, respectively, by 300-epoch training. The performance is slightly better than recent works of MoCo v3 and DINO which adopt DeiT as the backbone, but with much lighter tricks. More importantly, the general-purpose Swin Transformer backbone enables us to also evaluate the learnt representations on downstream tasks such as object detection and semantic segmentation, in contrast to a few recent approaches built on ViT/DeiT which only report linear evaluation results on ImageNet-1K due to ViT/DeiT not tamed for these dense prediction tasks. We hope our results can facilitate more comprehensive evaluation of self-supervised learning methods designed for Transformer architectures. Our code and models are available at https://github.com/SwinTransformer/Transformer-SSL, which will be continually enriched.
Multi-spectral frequency combs provide frontier architectures for laser standoff spectroscopy, optical clockwork, and high-capacity optical communications. Frequency microcombs, aided by their high-quality resonances and inherent third-order nonlinea
r susceptibility, have demonstrated remarkable impact in frequency metrology and synthesis. However, microcombs are often with limited spectral bandwidth bounded by the intrinsic second-order chromatic dispersion and the consequently low intensities at the spectral edges. To enhance the applications of frequency combs, a spectrally-broad comb generation scheme is often desired. Here we report coherent satellite clusters in multi-spectral regenerative frequency microcombs with enhanced intensities at the octave points and engineered frequency span. Beyond the conventional bandwidth of parametric oscillation, the regenerative satellites are facilitated by higher-order dispersion control allowing for multi-phase-matching in the microcavities. The frequency span of the multi-spectral regenerative satellites is deterministically controlled from 57 THz to 126 THz by pumping at C- and L- bands. We demonstrate that the regenerative satellite preserves the coherence with the central comb through the nonlinear parametric process. We further show the mirrored appearance of the satellite transition dynamics including each comb state that are simultaneously observed at the central comb. These multi-spectral regenerative satellites extend the scope of parametric-based frequency combs and provide a unique platform with wide applications.
Chi-3 micro resonators have enabled compact and portable frequency comb generation, but require sophisticated dispersion control. Here we demonstrate an alternative approach using a chi-2 sheet cavity, where the dispersion requirement is relaxed by c
avity phase matching. 21.2 THz broadband comb generation is achieved with uniform line spacing of 133.0 GHz, despite a relatively large dispersion of 275.4 fs^2/mm around 1064nm. With 22.6 % high slope efficiency and 14.9 kW peak power handling, this chi-2 comb can be further stabilized for navigation, telecommunication, astronomy, and spectroscopy applications.
Quantum entanglement is a fundamental resource for secure information processing and communications, where hyperentanglement or high-dimensional entanglement has been separately proposed towards high data capacity and error resilience. The continuous
-variable nature of the energy-time entanglement makes it an ideal candidate for efficient high-dimensional coding with minimal limitations. Here we demonstrate the first simultaneous high-dimensional hyperentanglement using a biphoton frequency comb to harness the full potential in both energy and time domain. The long-postulated Hong-Ou-Mandel quantum revival is exhibited, with up to 19 time-bins, 96.5% visibilities. We further witness the high-dimensional energy-time entanglement through Franson revivals, which is observed periodically at integer time-bins, with 97.8% visibility. This qudit state is observed to simultaneously violate the generalized Bell inequality by up to 10.95 deviations while observing recurrent Clauser-Horne-Shimony-Holt S-parameters up to 2.76. Our biphoton frequency comb provides a platform in photon-efficient quantum communications towards the ultimate channel capacity through energy-time-polarization high-dimensional encoding.
Near-infrared Hong-Ou-Mandel quantum interference is observed in silicon nanophotonic directional couplers with raw visibilities on-chip at 90.5%. Spectrally-bright 1557-nm two-photon states are generated in a periodically-poled KTiOPO4 waveguide chi
p, serving as the entangled photon source and pumped with a self-injection locked laser, for the photon statistical measurements. Efficient four-port coupling in the communications C-band and in the high-index-contrast silicon photonics platform is demonstrated, with matching theoretical predictions of the quantum interference visibility. Constituents for the residual quantum visibility imperfection are examined, supported with theoretical analysis of the sequentially-triggered multipair biphoton contribution and techniques for visibility compensation, towards scalable high-bitrate quantum information processing and communications.
We wish to report an experimental observation of anti-correlation from first-order incoherent classical chaotic light. We explain why the classical statistical theory does not apply and provide a quantum interpretation. In quantum theory, either corr
elation or anti-correlation is a two-photon interference phenomenon, which involves the superposition of two-photon amplitudes, a nonclassical entity corresponding to different yet indistinguishable alternative ways of producing a joint-photodetection event.
