LR0.FM benchmark evaluates the impact of low resolution on the zero-shot performance of VLMs, via novel Weighted Aggregated Robustness metric. LR-TK0 enhances the robustness of VLMs against low-resolution across several datasets without compromising their pre-trained weights.

Lightweight, annotation-free proxy for mitigating appearance bias in ReID models, when "expensive" clothing annotations aren't available. Colors See, Colors Ignore (CSCI), an efficient RGB-only method leverages color information directly from raw images or video frames.

We introduce Robustness against Low-Quality (RLQ) in clothes changing real-world ReID to make the model robust against low-quality artifacts like pixelation, out-of-focus blur, and motion blur.

One of the first works in Person Reid exploring multiple techniques to improve accuracy on the Video ReID model, and proposing novel attention loss for helping model focus on certain frames more than the other.

Removing bells and whistles of SOTA Zero-Shot Object Detectors under resolution degradation (`pixelation'), exposing 1) Vunerability of "backbones", 2) Certain images are immune to degradation 3) Minimal impact of "lanugage".

CosPlan benchmarks evaluates VLMs on "error-prone" sequence completion tasks, testing abilities like Error Detection and Sequence Completion ability. Novel SGI, incrementally updates Scene Graphs better handle decision making.

Layer Query Network (LQN), a lightweight five-layer MLP that adapts a frozen VLM in one forward pass. LQN distills intermediate-layer tokens from VLM and self-supervise achieving faster convergence and strong dense-prediction performance, outperforming the teacher VLM.