fsai-vision-utils 0.0.252

Helper functions for vision processing.

fsai-vision-utils

Vision utility functions and tools for batch processing and data management.

Installation

poetry add fsai-vision-utils

Tools

AWS Batch Download Tool

Download multiple files from S3 using file IDs with multi-threading and retry logic.

Usage

python -m fsai_vision_utils.clis.aws_batch_download \
    --ids_txt_file ids.txt \
    --aws_path s3://bucket/folder \
    --output_path ./images

Arguments

Argument	Required	Default	Description
--ids_txt_file	Yes	-	Text file with file IDs (one per line)
--aws_path	Yes	-	S3 base path (e.g., s3://bucket/folder)
--output_path	Yes	-	Local output directory
--file_extension	No	jpg	File extension to download
--max_workers	No	50	Number of concurrent downloads
--max_retries	No	3	Retry attempts per file
--log_level	No	INFO	Logging level

Input Format

Create a text file with one file ID per line:

image_001
image_002
image_003

The tool downloads: {aws_path}/{file_id}.{file_extension}

Examples

Basic download:

python -m fsai_vision_utils.clis.aws_batch_download \
    --ids_txt_file image_ids.txt \
    --aws_path s3://my-bucket/images \
    --output_path ./images

Custom settings:

python -m fsai_vision_utils.clis.aws_batch_download \
    --ids_txt_file image_ids.txt \
    --aws_path s3://my-bucket/images \
    --output_path ./images \
    --file_extension png \
    --max_workers 100

Features

• Multi-threaded downloads (configurable workers)
• Automatic retry with exponential backoff
• Skips already downloaded files
• Progress tracking and statistics
• Graceful shutdown (Ctrl+C)
• Comprehensive logging

Requirements

• AWS CLI installed and configured
• Valid AWS credentials with S3 read access

COCO Dataset Slicer

Slice COCO dataset images and annotations into smaller tiles for training or inference on large images.

Usage

python -m fsai_vision_utils.clis.coco_slice_dataset \
    --input-coco-json annotations.json \
    --image-dir ./images \
    --output-dir ./sliced_output \
    --slice-height 1024 \
    --slice-width 1024

Arguments

Argument	Required	Default	Description
--input-coco-json	Yes	-	Path to the input COCO annotations JSON file
--image-dir	Yes	-	Directory containing the input images
--output-dir	Yes	-	Directory to save the sliced images and annotations
--slice-height	No	1024	Height of each image slice in pixels
--slice-width	No	1024	Width of each image slice in pixels
--overlap-height-ratio	No	0.2	Overlap ratio for height between slices (0.0-1.0)
--overlap-width-ratio	No	0.2	Overlap ratio for width between slices (0.0-1.0)
--output-coco-json	No	output_dir/coco-annotations-sliced_coco.json	Path to save the output COCO JSON file

Input Format

The tool expects:

• A COCO format JSON file with annotations
• Images referenced in the COCO file located in the specified image directory
• Images can be in common formats (JPG, PNG, etc.)

Examples

Basic slicing with default settings:

python -m fsai_vision_utils.clis.coco_slice_dataset \
    --input-coco-json annotations.json \
    --image-dir ./images \
    --output-dir ./sliced_output

Custom slice size and overlap:

python -m fsai_vision_utils.clis.coco_slice_dataset \
    --input-coco-json annotations.json \
    --image-dir ./images \
    --output-dir ./sliced_output \
    --slice-height 512 \
    --slice-width 512 \
    --overlap-height-ratio 0.3 \
    --overlap-width-ratio 0.3

Specify custom output JSON path:

python -m fsai_vision_utils.clis.coco_slice_dataset \
    --input-coco-json annotations.json \
    --image-dir ./images \
    --output-dir ./sliced_output \
    --output-coco-json ./custom_output.json

Features

• Parallel Processing: Uses ProcessPoolExecutor for efficient multi-core processing
• Annotation Preservation: Maintains bounding box annotations across slices
• Overlap Control: Configurable overlap ratios to ensure no objects are missed
• Area Filtering: Automatically filters annotations based on minimum area ratio
• Error Handling: Gracefully handles invalid annotations and continues processing
• Progress Tracking: Shows completion status for each image
• Flexible Output: Supports custom output paths and formats

Output Structure

The tool creates:

• Sliced images in the output directory with naming pattern: {original_name}_{slice_index}.jpg
• A new COCO JSON file with updated annotations for all slices
• Preserved category information from the original dataset

Technical Details

• Minimum Area Ratio: Default 0.4 (40% of annotation must be visible in slice)
• File Extension: Output images are saved as JPG format
• Workers: Uses 8 parallel workers by default for optimal performance
• Memory Efficient: Loads images once per processing thread
• Topology Safe: Handles invalid polygon annotations gracefully

Use Cases

• Large Image Processing: Break down high-resolution images for model training
• Memory Optimization: Reduce memory requirements for inference
• Data Augmentation: Create overlapping slices for better model generalization
• Edge Case Handling: Ensure small objects aren't missed at slice boundaries