Hi authors,
Thank you so much for your fantastic work! I am a learner currently studying the PANDA framework, and I am very impressed by the idea of using an Agentic workflow for VAD.
I have been studying the qualitative visualization in Figure 4 (UCF-Crime) very carefully, but I find myself a bit confused about the exact timeline and the tool-invocation logic. Could you kindly help clarify a few details?
-
CCTV Timestamps Order: I noticed that the raw CCTV timestamps on the top two large images in the middle column don't seem to be strictly in chronological order (the first one appears to be 02:40:39, and the second one is 02:40:15). Could you explain how these frames are sequenced relative to the Video Time?
-
The "Insufficient" Judgment: In the second image (CCTV timestamp 02:40:15), the scene appears to be completely empty with no people present. Is it a bit counter-intuitive for the model to classify this seemingly empty scene as insufficient rather than normal? What triggers this uncertainty?
-
Image Retrieval Timeline & The Small Images: During Reflection Round 1, the model uses the Image Retrieve tool. However, the small images shown below have timestamps starting from 02:42:17, which seems to be in the future relative to the base frame of 02:40:15. How does the retrieval mechanism fetch these frames? Also, do these small images represent the specific frames retrieved by the tool, or do they represent the continuous forward streaming of the video?
-
Alignment of the Final "Abnormal" Score: After two rounds of reflection, the model successfully identifies the anomaly and outputs a score of 0.7 (abnormal). Which specific video frame is this "abnormal" label assigned to? Does it apply back to the original frame that triggered the reflection (e.g., 02:40:15), or is it assigned to the later frame where the person is clearly visible behind the counter (e.g., 02:43:11)?
Thank you so much for your time and patience! I really appreciate your help in helping me understand these details better.
Best regards.
Hi authors,
Thank you so much for your fantastic work! I am a learner currently studying the PANDA framework, and I am very impressed by the idea of using an Agentic workflow for VAD.
I have been studying the qualitative visualization in Figure 4 (UCF-Crime) very carefully, but I find myself a bit confused about the exact timeline and the tool-invocation logic. Could you kindly help clarify a few details?
CCTV Timestamps Order: I noticed that the raw CCTV timestamps on the top two large images in the middle column don't seem to be strictly in chronological order (the first one appears to be 02:40:39, and the second one is 02:40:15). Could you explain how these frames are sequenced relative to the Video Time?
The "Insufficient" Judgment: In the second image (CCTV timestamp 02:40:15), the scene appears to be completely empty with no people present. Is it a bit counter-intuitive for the model to classify this seemingly empty scene as insufficient rather than normal? What triggers this uncertainty?
Image Retrieval Timeline & The Small Images: During Reflection Round 1, the model uses the Image Retrieve tool. However, the small images shown below have timestamps starting from 02:42:17, which seems to be in the future relative to the base frame of 02:40:15. How does the retrieval mechanism fetch these frames? Also, do these small images represent the specific frames retrieved by the tool, or do they represent the continuous forward streaming of the video?
Alignment of the Final "Abnormal" Score: After two rounds of reflection, the model successfully identifies the anomaly and outputs a score of 0.7 (abnormal). Which specific video frame is this "abnormal" label assigned to? Does it apply back to the original frame that triggered the reflection (e.g., 02:40:15), or is it assigned to the later frame where the person is clearly visible behind the counter (e.g., 02:43:11)?
Thank you so much for your time and patience! I really appreciate your help in helping me understand these details better.
Best regards.