arXiv subjects

Hiroyuki Uchiyama

Publications and source records attributed to Hiroyuki Uchiyama.

3 recordsLinked to original sources

arXiv2024

Learning Multiple Representations with Inconsistency-Guided Detail Regularization for Mask-Guided Matting

Mask-guided matting networks have achieved significant improvements and have shown great potential in practical applications in recent years. However, simply learning matting representation from synthetic and lack-of-real-world-diversity matting data, these approaches tend to overfit low-level details in wrong regions, lack generalization to objects with complex structures and real-world scenes such as shadows, as well as suffer from interference of background lines or textures. To address these challenges, in this paper, we propose a novel auxiliary learning framework for mask-guided matting models, incorporating three auxiliary tasks: semantic segmentation, edge detection, and background line detection besides matting, to learn different and effective representations from different types of data and annotations. Our framework and model introduce the following key aspects: (1) to learn real-world adaptive semantic representation for objects with diverse and complex structures under real-world scenes, we introduce extra semantic segmentation and edge detection tasks on more diverse real-world data with segmentation annotations; (2) to avoid overfitting on low-level details, we propose a module to utilize the inconsistency between learned segmentation and matting representations to regularize detail refinement; (3) we propose a novel background line detection task into our auxiliary learning framework, to suppress interference of background lines or textures. In addition, we propose a high-quality matting benchmark, Plant-Mat, to evaluate matting methods on complex structures. Extensively quantitative and qualitative results show that our approach outperforms state-of-the-art mask-guided methods.

cs.CV↗

arXiv2023

Scheduling Space Expander: An Extension of Concurrency Control for Data Ingestion Queries

With the continuing advances of sensing devices and IoT/Telecom applications, database systems need to process data ingestion queries that update the sensor data frequently. However, as the rate of data ingestion queries increases, existing protocols have exhibited degraded performance since concurrent updates need to acquire lock to update the latest versions. To reduce the load on system on data ingestion queries, we focus on the theory of version order; we can test that a write is an old and unnecessary version by using version order of data items. In this paper, we propose a novel protocol extension method, scheduling space expander (SSE). SSE adds another control flow to conventional protocols to omit updates on data ingestion queries. It generates an erasing version order, which assumes that a transaction processes outdated unnecessary versions. SSE also tests the correctness of this version order efficiently and independently from conventional protocols. In addition, we present an optimization of SSE called epoch-based SSE (ESSE), which tests and maintains an erasing version order more efficiently than SSE. We extend two state-of-the-art 1VCC and MVCC protocols, Silo and MVTO with ESSE. Experimental results demonstrate that extensions of Silo and MVTO improve 2.7x and 2.5x performance on the TATP benchmark on a 144-core machine, and the extensions achieved performance comparable to that of the original protocol for the TPC-C benchmark.

cs.DB↗

arXiv2020

NWR: Rethinking Thomas Write Rule for Omittable Write Operations

Concurrency control protocols are the key to scaling current DBMS performances. They efficiently interleave read and write operations in transactions, but occasionally they restrict concurrency by using coordination such as exclusive lockings. Although exclusive lockings ensure the correctness of DBMS, it incurs serious performance penalties on multi-core environments. In particular, existing protocols generally suffer from emerging highly write contended workloads, since they use innumerable lockings for write operations. In this paper, we rethink the Thomas write rule (TWR), which allows the timestamp ordering (T/O) protocol to omit write operations without any lockings. We formalize the notion of omitting and decouple it from the T/O protocol implementation, in order to define a new rule named non-visible write rule (NWR). When the rules of NWR are satisfied, any protocol can in theory generate omittable write operations with preserving the correctness without any lockings. In the experiments, we implement three NWR-extended protocols: Silo+NWR, TicToc+NWR, and MVTO+NWR. Experimental results demonstrate the efficiency and the low-overhead property of the extended protocols. We confirm that NWR-extended protocols achieve more than 11x faster than the originals in the best case of highly write contended YCSB-A and comparable performance with the originals in the other workloads.

cs.DB↗