lower experimental accuracy

[LANGZHIZHEN]

Thank you for your valuable contribution to the field.
I downloaded the code and trained the multiviewx dataset. After running 10 epochs, I discovered that the recall and precision were extremely low. What could be the cause? Since there was only one GPU in my computer, I replaced cuda:1 with cuda:0.

[hou-yz]

Sorry for the late reply.

It seems from the results that the MODP [1] is very low, indicating that the localization accuracy (i.e., how close the predicted locations are compared to the ground truth) is not good enough.

Plus, the recall is also low. This means that within a radius of 0.5 meters, most ground truth pedestrian locations could not find a corresponding detection. Combined with the low MODP, it seems that the system is giving poor localization performance overall.

Can you please confirm that the evaluation is working correctly by running the following? It should give "eval: MODA 88.4, MODP 75.6, prec 93.6, rcll 95.0".

export PYTHONPATH=your/path/to/MVDet
cd your/path/to/MVDet/multiview_detector/evaluation
python evaluate.py

Hope this helps.

Best,
Yunzhong

[1]. Kasturi, Rangachar, Dmitry Goldgof, Padmanabhan Soundararajan, Vasant Manohar, John Garofolo, Rachel Bowers, Matthew Boonstra, Valentina Korzhova, and Jing Zhang. "Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and protocol." IEEE transactions on pattern analysis and machine intelligence 31, no. 2 (2008): 319-336.