Skip to main content
Not all annotation formats carry the same information. When you convert between formats, some fields may not have an equivalent in the target format — that’s what Panlabel calls “lossiness.” Rather than silently dropping data, Panlabel tells you exactly what would be lost and asks you to opt in with --allow-lossy. Every convert command generates a report explaining what happened.

​
Lossiness Model

A conversion report issue has a severity:
  • warning: real information loss risk; conversion is blocked unless --allow-lossy is set
  • info: deterministic policy note; never blocks conversion

​
Format-Level Lossiness

Each format has a general lossiness classification relative to Panlabel’s intermediate representation (IR):
FormatLossinessDescription
ir-jsonLosslessCan represent everything in the IR (round-trip safe)
cocoConditionalMay lose some information depending on dataset content
cvatLossyAlways loses some IR information
label-studioLossyAlways loses some IR information
tfodLossyAlways loses some IR information
yoloLossyAlways loses some IR information
vocLossyAlways loses some IR information
The format-level class is a general capability signal. Conversions are actually blocked only when the report contains one or more warning issues.

​
JSON Report Format

With --report json, the output follows this shape: 
{
  "from": "coco",
  "to": "yolo",
  "input": {"images": 10, "categories": 3, "annotations": 40},
  "output": {"images": 10, "categories": 3, "annotations": 40},
  "issues": [
    {"severity": "warning", "code": "drop_dataset_info", "message": "..."},
    {"severity": "info", "code": "yolo_writer_float_precision", "message": "..."}
  ]
}
The report includes:
  • from/to: source and target format names
  • input/output: counts of images, categories, and annotations before and after conversion
  • issues: list of warnings and info notes with stable codes

​
Stable Issue Codes

These codes are designed to be stable for programmatic use.

​
Warning Codes

CodeMeaning
drop_dataset_infoDataset-level metadata is dropped
drop_licensesLicense list is dropped
drop_image_metadataImage metadata fields (license/date) are dropped
drop_category_supercategoryCategory supercategory is dropped
drop_annotation_confidenceAnnotation confidence values are dropped
drop_annotation_attributesAnnotation attributes are dropped
drop_images_without_annotationsImages without annotations will not appear (TFOD behavior)
drop_dataset_info_nameinfo.name has no COCO equivalent
coco_attributes_may_not_be_preservedSome COCO-tool roundtrips may not preserve nonstandard attributes
label_studio_rotation_droppedRotated Label Studio boxes are flattened to axis-aligned envelopes; angle is kept as ls_rotation_deg attribute

​
Info Codes

CodeMeaning
tfod_reader_id_assignmentTFOD reader deterministic ID policy
tfod_writer_row_orderTFOD writer deterministic row order
yolo_reader_id_assignmentYOLO reader deterministic ID policy
yolo_reader_class_map_sourceYOLO class map precedence/source note
yolo_writer_class_orderYOLO writer class index assignment policy
yolo_writer_empty_label_filesYOLO writer creates empty label files for unannotated images
yolo_writer_float_precisionYOLO normalized float precision policy
voc_reader_id_assignmentVOC reader deterministic ID assignment policy
voc_reader_attribute_mappingVOC reader mapping of pose/truncated/difficult/occluded attributes
voc_reader_coordinate_policyVOC reader coordinate policy (no 0/1-based adjustment)
voc_reader_depth_handlingVOC reader depth metadata handling note
voc_writer_file_layoutVOC writer XML path/layout policy
voc_writer_no_image_copyVOC writer placeholder JPEGImages policy
voc_writer_bool_normalizationVOC writer boolean normalization policy
label_studio_reader_id_assignmentLabel Studio reader deterministic ID assignment policy
label_studio_reader_image_ref_policyLabel Studio reader image reference mapping policy
label_studio_writer_from_to_defaultsLabel Studio writer default from_name / to_name policy
cvat_reader_id_assignmentCVAT reader deterministic ID assignment policy
cvat_reader_attribute_policyCVAT reader coordinate + attribute mapping policy
cvat_writer_meta_defaultsCVAT writer minimal <meta> block policy

​
Examples of Lossy Conversions

​
COCO to YOLO

When converting from COCO to YOLO, several types of information are lost: 
panlabel convert -f coco -t yolo -i dataset.json -o ./yolo_out --allow-lossy
This conversion will drop:
  • Dataset-level metadata (info fields)
  • License information
  • Image metadata (license_id, date_captured)
  • Category supercategories
  • Annotation confidence scores
  • Annotation attributes

​
TFOD Behavior

TFOD has a unique behavior: images without annotations will not appear in the output. 
panlabel convert -f coco -t tfod -i dataset.json -o output.csv --allow-lossy
If your input has 100 images but only 80 have annotations, the output will contain only those 80 images.

​
VOC Attribute Preservation

VOC can preserve some annotation attributes but drops others: 
panlabel convert -f coco -t voc -i dataset.json -o ./voc_out --allow-lossy
Preserved attributes:
  • pose
  • truncated
  • difficult
  • occluded
All other custom attributes will be dropped.

​
Practical Guidance

1

Check the conversion report

If conversion is blocked, run once with --report json to see exactly which warning codes triggered the block:
panlabel convert -f coco -t yolo -i input.json -o output --report json
2

Review what will be lost

Examine the issues array in the JSON report to understand what information will be dropped.
3

Opt in with --allow-lossy

Only use --allow-lossy when you accept those specific losses:
panlabel convert -f coco -t yolo -i input.json -o output --allow-lossy
Prefer explicit --allow-lossy only when you accept those specific losses. Never use it blindly without reviewing the conversion report first.

​
Understanding Conditional Lossiness

The COCO format is marked as “Conditional” because it may lose some information depending on your dataset content:
  • Dataset name (info.name): COCO has no equivalent field, so this is always lost
  • Custom attributes: COCO round-trips area and iscrowd via attributes, but other attributes may not be preserved by COCO tools
# Converting IR JSON to COCO may trigger warnings
panlabel convert -f ir-json -t coco -i dataset.json -o coco_output.json
If your IR dataset has a name field or custom attributes, you’ll see warnings about potential data loss.

Build docs developers (and LLMs) love

Get started for freeTalk to us