← Back to case studies Synthetic vision data case

Port and gantry vision data simulation: image augmentation, synthetic scenes, and dataset quality control

An anonymized Sinovanta case covering port-scene source auditing, eight-condition image augmentation, generated object scenes, metadata, quality checks, and review documents.

Engagement
Source-data audit and synthetic image demonstration
Scope
Port and yard scenes, image augmentation, generated samples, quality checks, and documentation
Input boundary
Image archive and four-class list; no bounding boxes or segmentation masks
Verified outputs
48 augmentation samples, 40 generated class images, and 3 pose/background examples
01

Project brief

The source archive contained port and gantry-related images plus a four-class list: forklift, cone, bollard, and person. It did not contain detection boxes or segmentation masks. The task was to establish a controlled demonstration set for weather, lighting, camera degradation, object presentation, and document review before any model-training decision.

02

Delivery objective

Build a traceable package that separates source-derived augmentation from generated visual examples. Every transformed image needed a scenario, source reference, seed, and parameter record. The result was intended for dataset planning and review, not as proof of model accuracy or as annotation-ready production data.

Implementation workflow

The workflow combined source governance, reproducible image processing, generated scene examples, quantitative file checks, and visual review.

  1. 01

    Audit the source archive

    Inventory image files, dimensions, class definitions, and available annotation formats. Record that the supplied archive had a class list but no boxes or masks.

  2. 02

    Control source eligibility

    Review filenames and contact sheets, exclude unrelated external road imagery, and regenerate the package from the accepted port and yard source set.

  3. 03

    Generate eight stress conditions

    Create six 1280 x 720 samples for each of clear day, night yard, rain camera, fog or haze, sunset backlight, motion blur, foreground occlusion, and surveillance compression.

  4. 04

    Create class-oriented scene examples

    Produce ten generated images for each of four review classes: crane, container, cone, and forklift. Keep these files separate from source-derived augmentation.

  5. 05

    Review pose and background variation

    Create three additional examples covering city-road curb tilt, construction-site lean, and factory yaw or tilt to support composition discussion.

  6. 06

    Package evidence and documentation

    Write metadata and quality summaries, render two contact sheets, and prepare a 19-page customer document containing 16 representative large-format images.

03

Delivered package

Reproducible generator
A parameterized script for the eight image-level stress conditions, with stable seeds and output metadata.
48-image augmentation set
Eight scenarios with six 1280 x 720 JPG files per scenario and one metadata row for every output.
40 generated class images
Ten visual examples for each of crane, container, cone, and forklift, stored separately from source-derived images.
Pose and background examples
Three generated examples used to review object orientation, framing, and environmental variation.
Customer review document
A 19-page DOCX and PDF presentation with 16 selected images, package scope, and use boundaries.
04

Verified evidence

The following figures are file, metadata, and document checks from the delivered package. They are not model-performance metrics.

Source audit

Four class names; no bounding boxes or masks.

The available evidence supported image-level simulation and planning only. It did not support supervised detection training without additional annotation.

Augmentation output

48 JPG files at 1280 x 720 across 8 scenarios.

Each scenario contains six files. Metadata records the source image, scenario, random seed, and transformation parameters.

Quality checks

0 errors, unknown rows, external-source rows, low-information candidates, or near-duplicate pairs.

Measured brightness ranged from 39.4 to 137.5 and image standard deviation from 15.5 to 62.6 across the 48-file set.

Generated review assets

40 class images, 3 pose/background examples, and a 19-page document with 16 selected images.

These counts confirm package completeness and document layout. They do not establish annotation quality or model suitability.

05

Validation performed

  • All 48 augmented files opened successfully and reported the expected 1280 x 720 dimensions.
  • The metadata table contained 48 rows, with six outputs assigned to each of the eight scenarios.
  • Both contact-sheet pages were reviewed after source filtering; no unrelated external road image remained in the final package.
  • The 40 generated class files and three pose/background files were counted and opened; no label-accuracy or model-accuracy test was performed.
06

Use and evidence boundary

This case demonstrates data auditing, image-level augmentation, generated visual examples, metadata, quality checks, and customer documentation. It does not claim that the source archive was annotation-ready, that the generated images are verified training labels, or that any detector reached a stated precision, recall, or mAP. Before model development, the dataset still requires annotation, train/validation partitioning, real-scene representativeness review, class-balance analysis, annotation quality assurance, and metric-based acceptance criteria.

Planning a vision dataset for port or industrial scenes?

Share the source structure, target classes, operating conditions, annotation status, and intended model task. Sinovanta can define a reviewable data and validation scope.

Discuss the data scope →