ChiViT
ChiViT is a band-agnostic vision transformer pretrained for remote-sensing representation learning and evaluated as a supporting baseline in GeoCrossBench.
Checkpoint
- Repository:
yerevann/ChiViT - Pinned revision:
dad320e530f45964a03b2a4c863e65217a989dfd - File:
checkpoint.pth - Size: 1,539,025,955 bytes
- SHA-256:
0fd0913d069bc236183c1bcb025c8319f62578af45fc814aeb8d9a828bb4ef6f
The checksum matches the checkpoint used by the recovered GeoCrossBench evaluator.
Training summary
The saved checkpoint metadata records a 400,000,000-sample budget, 781,200 completed optimizer iterations, batch size 16 per GPU, eight workers, and gradient accumulation of four. This corresponds to 399,974,400 processed samples. The model used self-supervised pretraining with band sampling over remote-sensing imagery, including paired optical and SAR data.
The standalone repository launcher specifies 500M samples and is not the final run configuration. See chivit-training.json for the checkpoint-derived record.
Evaluation and limitations
ChiViT was evaluated on GeoCrossBench across in-distribution, no-overlap, and superset band-transfer settings. Benchmark results mix task-specific metrics; consult protocol.json before interpreting aggregate scores.
A post-hoc audit found that nearly all images in the GeoBench BigEarthNet test split appeared without labels in ChiViT's self-supervised pretraining corpus. BigEarthNet labels were not used for pretraining. Report this overlap whenever publishing BigEarthNet results and avoid treating that task as a clean unseen-data test.
The released checkpoint does not establish performance on future sensors or satellites beyond the band and modality shifts measured by GeoCrossBench.
Intended use
Use this checkpoint for research on remote-sensing representation learning and cross-band transfer. Downstream users must validate sensor calibration, preprocessing, spatial alignment, and task suitability for their own data.
License
The ChiViT checkpoint is released under the Apache License 2.0. See LICENSE in the model repository.