update talk

_pages/WACV_2026.md

@@ -76,13 +76,25 @@ TBD
 
 ### Keynote Speakers
 <div class="row projects pt-1 pb-1">
+      <div class="col-sm-4">
+          {% include people.html name="Tong Shen" affiliation="CMU" img="assets/img/TongShen.jpg" %}
+      </div>
       <div class="col-sm-4">
           {% include people.html name="Litian Liu" affiliation="Qualcomm" url="https://litianliu.github.io/" img="https://litianliu.github.io/assets/img/prof_pic.jpg?0f86e7255c77a6dfd89f051d80803a8d" %}
       </div>
 </div>
 
 ### Keynote Talks
 
+#### Teaching Trackers to Think: From Appearance Matching to Agentic Perception (Tong Shen)
+**Abstract:**
+Deep learning trackers have made remarkable progress on standard benchmarks, yet they still fail catastrophically at critical moments, such as occlusions, near visually similar distractors, or through rapid motion. Humans, by contrast, track effortlessly through these challenges. We don't just match appearances; we reason about identity ("that rabbit has no black markings"), predict through occlusion ("it should reappear on the other side"), and recognize our own mistakes ("wait, I'm following the wrong one"). We hypothesize that this cognitive dimension, world knowledge and reasoning beyond what any tracking dataset can provide, is what makes human tracking robust, and that bridging this gap is the key to approaching human-level performance.
+
+This talk presents a research arc toward closing that intelligence gap. We show how vision-language models can serve as a sparse cognitive layer that intervenes at critical decision points, using contrastive semantic reasoning to disambiguate targets from distractors. We then explore how this reactive intervention can evolve into genuine self-awareness: trackers that watch replays of their own behavior, diagnose failures, and self-correct. We also argue that this new paradigm demands new evaluation, and present a benchmark measuring tracker decision-making during challenging events and reveal the diagnostic insights hidden by conventional benchmarks. Looking ahead, we envision the next generation tracker to be foundation models that perceive, track, and reason about the visual world as a single integrated process.
+
+**Bio:**
+Tong Shen is a PhD candidate in the Department of Electrical and Computer Engineering at Carnegie Mellon University, advised by José M. F. Moura. His current research focuses on agentic video understanding, combining large vision-language models with perception and reasoning to build robust real-world systems. He is broadly interested in machine perception, autonomous decision-making, and planning for robotics, with an emphasis on bridging foundation models with reliable, goal-directed behavior. He received his M.S. from Purdue University and his B.S. from Xi'an Jiaotong University. His prior work spans unsupervised domain adaptation for object detection, latency-aware DNN optimization, and spatiotemporal modeling for COVID-19 forecasting.
+
 #### Towards robust and efficient VLA for end-to-end autonomous driving (Litian Liu)
 
 **Abstract:**