Hi, I would love to be added to it if possible. I am a PhD student at UIUC working on vision-language models.

(8/8)
Paper: openaccess.thecvf.com/content/WACV...

Project Page: samyakr99.github.io/PositiveCoOp/

#WACV2025 #AI #MachineLearning #ComputerVision #CLIP #MultiLabelRecognition #PromptLearning

(7/8) This work is done at UIUC with
@shubhangb.bsky.social and Prof. Narendra Ahuja

Excited to discuss more at WACV 2025! Come find us at Poster Session 3 - 2nd March 11:15-1PM

(6/8) TL;DR: If you're using VLMs for MLR, skip negative prompts and use learned embeddings instead!
This saves compute, parameters, and improves performance.

(5/8) Why is Negative Prompting Ineffective?
🔍 We analyze the LAION-400M dataset and find that less than 0.5% of captions contain negative words.
❌ CLIP simply doesn’t learn meaningful representations for class absence!

(4/8)Results on COCO & VOC2007
✅ PositiveCoOp outperforms existing dual-prompt methods (like DualCoOp)
✅ A simple vision-only baseline performs surprisingly well shows prompting isn’t always necessary!
✅ NegativeCoOp performs the worst, proves negative prompting is not optimal

(3/8) We introduce PositiveCoOp and NegativeCoOp:
🔹 PositiveCoOp learns only positive prompts via CLIP and replaces negative prompts with learned embeddings
🔹 NegativeCoOp does the opposite.
🔹 Which one works better? (Spoiler: PositiveCoOp wins! 🏆)

(2/8) We show that negative prompts hurt MLR performance:
👉 VLMs like CLIP are trained on image-caption data that focus on what’s present, not what’s absent.
👉 As a result, negative prompts often highlight the same regions as positive ones!

(1/8)Vision-language models like CLIP have been used for multi-label recognition (MLR) by learning both positive and negative prompts for associated with presence and absence of each class.
But is learning negative prompts actually helping detect absence? 🤔