AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

Robust Calibration of Large Vision-Language Adapters

Balamurali Murugesan, Julio Silva-Rodríguez, Ismail Ben Ayed, Jose Dolz

ECCV 2024 · Vision-language and multimodal › CLIP and zero-shot · approach: Test-time adaptation of VLMs · 14 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
adapt CLIP with Adapters, prompt learning or TPT instead of zero-shotCLIP · ImageNet OOD avgECE: 7.18 -> 15.45Hurt
apply sample-adaptive logit scaling (SaLS) at inferenceCLIP-Adapter (RN50) · ImageNet OOD avgECE: 15.45 -> 8.95Improved
apply SaLS at inference with no accuracy changeTIP-Adapter(f) (RN50) · ImageNet OOD avgECE: 19.04 -> 8.13Improved
add penalty constraint on logits outside zero-shot range during trainingTIP-Adapter(f) (ViT-B/16) · ImageNet OOD avgECE: 63.63 -> 40.98Improved
add penalty constraint on logits outside zero-shot range during trainingTIP-Adapter(f) (ViT-B/16) · ImageNet OOD avgaccuracy: 25.86 -> 49.23Improved
zero-shot logit normalization during training (ZS-Norm) (ablation)CLIP-Adapter (RN50) · ImageNet OOD avgECE: 15.45 -> 21.27Hurt
zero-shot logit normalization during training (ZS-Norm) (ablation)CLIP-Adapter (RN50) · ImageNet OOD avgaccuracy: 34.07 -> 30.06Hurt
apply SaLS to MaPLeMaPLe (ViT-B/16) · ImageNet OOD avgECE: 4.13 -> 4.38Hurt
apply calibration methods in test-time prompt tuningTPT (RN50) · 11 fine-grained datasetsECE: 11.27 -> 9.26Improved
apply calibration methods to TPTTPT (RN50) · 11 fine-grained datasetsaccuracy: 58.03 -> 58.03No clear effect
reduce zero-shot logit range to half or one quarter in SaLS (ablation)CoOp (RN50) · ImageNet shiftsECE: 7.82 -> 37.72Hurt
apply penalty constraint to CoOpCoOp (RN50) · ImageNet OOD avgaccuracy: 40.86 -> 41.87Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy