Paper Update

README.md

@@ -34,7 +34,7 @@ size_categories:
 
 <div align="center">
 
-<h1><a style="color:blue" href="https://daniel-cores.github.io/tvbench/">TVBench: Redesigning Video-Language Evaluation</a></h1>
+<h1><a style="color:blue" href="https://daniel-cores.github.io/tvbench/">Lost in Time: A New Temporal Benchmark for Video LLMs</a></h1>
 
 [Daniel Cores](https://scholar.google.com/citations?user=pJqkUWgAAAAJ)\*,
 [Michael Dorkenwald](https://scholar.google.com/citations?user=KY5nvLUAAAAJ)\*,
@@ -59,9 +59,18 @@ TVBench is a new benchmark specifically created to evaluate temporal understandi
 
 We defined 10 temporally challenging tasks that either require repetition counting (Action Count), properties about moving objects (Object Shuffle, Object Count, Moving Direction), temporal localization (Action Localization, Unexpected Action), temporal sequential ordering (Action Sequence, Scene Transition, Egocentric Sequence) and distinguishing between temporally hard Action Antonyms such as "Standing up" and "Sitting down".
 
-In TVBench, state-of-the-art text-only, image-based, and most video-language models perform close to random chance, with only the latest strong temporal models, such as Tarsier, outperforming the random baseline. In contrast to MVBench, the performance of these temporal models significantly drops when videos are reversed.
+In TVBench, state-of-the-art text-only, image-based, and most video-language models perform close to random chance, with only the latest strong temporal models, such as Tarsier, outperforming the random baseline.
+
+<center>
+<img src="figs/fig1.png" alt="drawing" width="600"/>
+</center>
+
+The performance of a SOTA models such as Tarsier on commonly used benchmarks hardly drops when shuffling the input videos. This suggests that these benchmarks do not effectively measure temporal understanding. In contrast, in our proposed TVBench, shuffling input frames results in random accuracy, as it should be.
+
+<center>
+<img src="figs/benchmarks.png" alt="drawing" width="600"/>
+</center>
 
-![image](figs/fig1.png)
 
 ### Dataset statistics:
 The table below shows the number of samples and the average frame length for each task in TVBench.
@@ -84,7 +93,7 @@ If you find this benchmark useful, please consider citing:
 
 @misc{cores2024tvbench,
   author = {Daniel Cores and Michael Dorkenwald and Manuel Mucientes and Cees G. M. Snoek and Yuki M. Asano},
-  title = {TVBench: Redesigning Video-Language Evaluation},
+  title = {Lost in Time: A New Temporal Benchmark for Video LLMs},
   year = {2024},
   eprint = {arXiv:2410.07752},
 }