Uniloc USA Inc v. Amazon.com Inc

I. Executive Summary and Procedural Information

• Parties & Counsel:
  • Plaintiff: Uniloc USA, Inc. (Texas); Uniloc Licensing USA LLC (Delaware); Uniloc 2017, LLC (Delaware)
  • Defendant: Amazon.com, Inc. and related entities (Delaware)
  • Plaintiff’s Counsel: Etheridge Law Group, PLLC
• Case Identification: 2:18-cv-00379, E.D. Tex., 07/31/2018
• Venue Allegations: Plaintiff alleges venue is proper in the Eastern District of Texas because Amazon is registered to do business in Texas, has committed acts of infringement in the district, and maintains a regular and established place of business there, including distribution facilities and data centers.
• Core Dispute: Plaintiff alleges that Defendant’s Amazon Web Services (AWS) platform, specifically its Rekognition and Sagemaker services, infringes a patent related to methods for scalable, content-based image retrieval.
• Technical Context: The technology concerns systems for efficiently searching large databases of images by analyzing their visual content, rather than relying on manual text-based tags, a foundational capability for modern AI-driven image analysis platforms.
• Key Procedural History: The complaint does not mention any prior litigation, Inter Partes Review (IPR) proceedings, or licensing history related to the patent-in-suit.

Case Timeline

II. Technology and Patent(s)-in-Suit Analysis

U.S. Patent No. 6,253,201 - "SCALABLE SOLUTION FOR IMAGE RETRIEVAL"

• Patent Identification: U.S. Patent No. 6,253,201, "SCALABLE SOLUTION FOR IMAGE RETRIEVAL," issued June 26, 2001.

The Invention Explained

• Problem Addressed: The patent describes the challenge of retrieving images from large databases. It notes that text-based keyword searching is subjective and burdensome, while graphics-based systems that compare images pixel-by-pixel or by a single global characteristic (e.g., overall color composition) are computationally impractical and can produce irrelevant results as databases grow ('201 Patent, col. 1:10-2:8).
• The Patented Solution: The invention proposes a method to avoid direct, computationally expensive image-to-image comparisons. An image is first broken down into a plurality of "content independent partitions" (e.g., a fixed grid). Each partition is then analyzed to create a descriptive "index value" based on its characteristics, such as color or edge patterns. The system maintains indexed lists that associate these index values with the identifiers of images containing a partition with that characteristic. To find a similar image, the system analyzes a target image's partitions, retrieves the corresponding lists of identifiers, and identifies which image identifiers appear most frequently, indicating a high degree of similarity without comparing the full images ('201 Patent, Abstract; col. 2:40-51).
• Technical Importance: This indexing method was designed to provide a scalable solution for content-based image retrieval that does not suffer from progressive performance degradation as the size of the image database increases ('201 Patent, col. 2:13-17).

Key Claims at a Glance

• The complaint asserts "one or more claims... including at least claim 5" (Compl. ¶25).
• Independent Claim 5 recites a method of indexing an image with the following essential elements:
  • identifying the image by an image identifier,
  • partitioning the image into a plurality of content independent partitions,
  • characterizing each content independent partition to form at least one index value, and
  • appending the image identifier to at least one list of image identifiers, where the list is determined by the index value that characterizes the partition.
• The complaint also makes infringement allegations under the doctrine of equivalents (Compl. ¶26).

III. The Accused Instrumentality

Product Identification

The "Accused Infringing Devices" are identified as the "Amazon Web Services" (AWS) platform, and specifically services and platforms including "Amazon Elastic Rekognition" ("Rekognition") and "Amazon Sagemaker" ("Sagemaker") (Compl. ¶18).

Functionality and Market Context

The complaint describes Rekognition as a visual analysis service that allows users to search and organize large image libraries. It allegedly works by "generating labels based on image analysis," which are then "indexed" (Compl. ¶20).

Sagemaker is described as a platform for building and deploying machine learning models, including models for object detection and classification in images using frameworks like the Single Shot multibox Detector (SSD) (Compl. ¶22). Sagemaker is also alleged to allow users to create a "searchable image library" by generating and indexing labels based on image analysis (Compl. ¶24).

The complaint alleges these services partition images into "bounding boxes or region of interests," generate "classes (index values)" for each partition, and enable indexing on these values using tools like Elasticsearch (Compl. ¶¶20, 24). No probative visual evidence provided in complaint.

IV. Analysis of Infringement Allegations

’201 Patent Infringement Allegations

Identified Points of Contention

• Scope Questions: A central dispute may arise over the term "content independent partitions." The patent specification's primary embodiment describes fixed grids (e.g., "a 4x4, 8x8, or 16x16 partitioning") applied uniformly to images ('201 Patent, col. 4:30-32). The complaint alleges infringement by systems that generate "bounding boxes or region of interests" (Compl. ¶20), which in modern AI are often content-dependent (i.e., drawn around a detected object). This raises the question of whether the claim term can be construed to read on such dynamically generated regions, even if the complaint asserts they are "independent of the content of the image" (Compl. ¶23).
• Technical Questions: The complaint equates the "labels/classes" generated by the accused AI services with the "index value" of the patent (Compl. ¶¶21, 24). The patent describes index values as quantized representations of low-level features like color or edges ('201 Patent, col. 4:36-40, FIG. 5). A technical question is whether a high-level semantic label (e.g., "car") produced by a neural network is the same as the quantized feature vector described in the patent.

V. Key Claim Terms for Construction

"content independent partitions"

• Context and Importance: The construction of this term is critical. If defined narrowly as only pre-determined, fixed grids, it may not cover the alleged "bounding boxes" or "regions of interest" used in modern object detection systems. Practitioners may focus on this term as it goes to the core of whether a 2001-era patent can read on modern AI techniques.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: The claim language itself does not specify a "fixed grid," only that the partitions are "content independent." A party could argue that any partitioning scheme applied algorithmically without full semantic understanding of the image meets this limitation. The complaint attempts to frame the accused functionality this way (Compl. ¶23).
  • Evidence for a Narrower Interpretation: The specification’s only concrete examples describe fixed geometric arrays: "typically the array is a 4x4, 8x8, or 16x16 partitioning of the image" ('201 Patent, col. 4:30-32). This language may be used to argue that the invention is limited to such non-adaptive partitioning structures.

"index value"

• Context and Importance: The infringement theory depends on equating the "labels" and "classes" from AWS with the claimed "index value." The definition will determine if a high-level, semantic classification from a machine learning model is equivalent to the patent's concept of a quantized low-level feature.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: The patent defines the term broadly as characterizing a partition as "one of a set of predefined indexed characterizations" ('201 Patent, col. 4:33-35). This could arguably encompass a text "label" or a numerical "class ID" from a predefined set.
  • Evidence for a Narrower Interpretation: The detailed description and figures primarily teach creating an index value by processing low-level visual data, such as a "histogram of the occurrences of the components of the descriptive characteristic" (e.g., colors or edge types), which is then quantized ('201 Patent, col. 4:36-40; FIG. 5). This could support an argument that an "index value" must be a direct quantization of low-level features, not a high-level semantic abstraction.

VI. Other Allegations

• Indirect Infringement: The complaint does not plead a separate count for indirect infringement and does not allege specific facts to support the knowledge and intent elements required for induced infringement, or the specific elements of contributory infringement. It mentions "indirect infringement" only in passing within its venue allegations (Compl. ¶13).
• Willful Infringement: Willful infringement is not alleged in the complaint.

VII. Analyst’s Conclusion: Key Questions for the Case

• A core issue will be one of definitional scope: can the term "content independent partitions", disclosed in the patent as a fixed grid, be construed to cover the algorithmically generated "bounding boxes" and "region of interests" that are fundamental to the accused modern object-detection services?
• A key question of technical interpretation will be whether the high-level semantic "labels" and "classes" generated by the accused AI platforms are equivalent to the "index value" claimed in the patent, which the specification describes as a quantized representation of low-level visual features like color and edges.
• An evidentiary question will center on proof of infringement: what evidence will show that the accused systems, which use general-purpose tools like Elasticsearch, perform the specific claimed step of "appending the image identifier to... lists of image identifiers" determined by an index value, as distinct from a more complex database operation that achieves a similar result?