Military-Labeled CLIP Crops: 4,844 Object Images from DVIDS

4,844 per-object image crops extracted from DVIDS military imagery, each accompanied by a Gemini-VLM caption. The dataset was created by llama-farm and last updated on 2026-05-29. It contains 12 object classes and is intended for CLIP fine-tuning or zero-shot evaluation.

Use Cases

• Fine-tuning CLIP models for zero-shot military object recognition based on the provided image crops and captions.
• Evaluating vision-language model performance on defense-related imagery based on the DVIDS-sourced objects.
• Training or benchmarking object detection models using the provided bounding box coordinates and class labels.
• Studying domain adaptation of pre-trained models to specialized military visual contexts.

Strengths

• Contains 4,844 labeled object crops, providing a focused collection for model training.
• Includes per-crop metadata such as bounding boxes, class IDs, and Gemini-VLM generated captions.
• Covers 12 distinct object classes relevant to military imagery.

Limitations

• Row count for the metadata file is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect geographic, temporal, or source bias inherent to the DVIDS platform.

Provenance

Source

DVIDS (Defense Visual Information Distribution Service)

Collection Method

Per-object crops extracted from DVIDS military imagery, with captions generated by Gemini-VLM.

Freshness

Last updated 2026-05-29 18:38:59; freshness should be verified.