📢 Announcement

VISIONxFAIL spotlights negative results, failure analysis, and real‑world constraints across CV/Graphics/Multimodal systems. We promote a culture of transparent reporting to accelerate scientific progress.

Website: lc4workshop.github.io/visionxfail.html • Estimated attendees: 20–30

🎯 Overview

Despite benchmark gains, many methods underperform in practice due to dataset biases, domain shifts, missing modalities, annotation noise, or unrealistic assumptions—yet such negative results are rarely documented. This workshop emphasizes: (1) understanding how constraints (data scarcity, supervision limits, compute bottlenecks) affect pipelines; (2) highlighting negative results in generative AI and vision systems; (3) documenting failure cases in cross‑domain, cross‑lingual, and cross‑modal applications; and (4) promoting transparent reporting of successes and failures.

Guiding Questions:
  • What constraints prevent vision/graphics systems from scaling robustly?
  • Which transfer/adaptation strategies have failed in real deployments?
  • How do cross‑modal alignments break under noise, missing data, or adversarial conditions?
  • How can documenting negative results prevent repeated mistakes?
  • Can we design evaluation frameworks that reward honest reporting of failures?

🧩 Topics of Interest

  • Negative results in detection, segmentation, tracking, recognition.
  • Failure analysis of ViTs, diffusion, GANs, and multimodal foundation models.
  • Constraints in cross‑modal & cross‑lingual tasks: OCR, scene text, multilingual captioning.
  • Understanding gaps and results in font processing and analysis.
  • Challenges in video understanding: action recognition, scene graphs, multi‑object tracking under constraints.
  • Annotation bottlenecks: bias, low inter‑rater reliability, annotation drift.
  • Graphics negative results: rendering, simulation, visual realism under limited resources.
  • Metrics that fail to capture robustness, fairness, or completeness.
  • Cross‑domain failures: adaptation/generalization gaps across datasets, genres, modalities.
  • Agentic workflows that failed due to latency, supervision gaps, or domain shifts.
  • Bias, fairness, and failure amplification in low‑resource vision settings.

Invited Speakers: TBA

🗓️ Tentative Program (Half‑day)

09:00 – 09:30Opening remarks & workshop intro
09:30 – 10:30Keynote / Invited Talk 1
10:30 – 11:00Coffee / Networking Break
11:00 – 12:30Contributed papers (6–8 papers, 12 min each)
12:30 – 14:00Lunch Break

Mode: In‑person planned; remote talks allowed. If the main conference goes fully virtual, the workshop will be virtual‑only.

📝 Submission & Review

  • Submission type: As per main conference guidelines.
  • Review process: Double‑blind with at least 2–3 reviews per submission.
  • Criteria: Relevance to theme, technical novelty/soundness, clarity, and potential for impact/discussion.
  • Management: Reviews coordinated via EasyChair (or equivalent) in alignment with main conference standards.

📚 Selected References

  1. Borji (2017). Negative Results in Computer Vision: A Perspective. arXiv:1705.04402.
  2. Zhang et al. (2014). Predicting Failures of Vision Systems. CVPR.
  3. Northcutt et al. (2021). Pervasive Label Errors Destabilize ML Benchmarks. arXiv:2103.14749.
  4. Hosseini et al. (2017). CNNs Are Not Invariant to Negative Images. arXiv:1703.06857.
  5. Lee et al. (2017). Towards Qualitative Analysis of ImageNet Classification Failures. arXiv:1709.03439.
  6. Zendel et al. (2017). Analyzing CV Datasets via Subspace Alignment & HAZOP. CVPR.
  7. Balayn et al. (2023). Handling Failures in DL‑based CV Systems. TiiS 13(2).
  8. Torralba & Efros (2011). Unbiased Look at Dataset Bias. CVPR.
  9. Recht et al. (2019). Do ImageNet Classifiers Generalize to ImageNet? ICML.
  10. Geirhos et al. (2018). Generalisation in Deep Learning. ICLR.