Permalink
Browse files

Changed README

  • Loading branch information...
Abel Gonzalez
Abel Gonzalez committed May 17, 2017
1 parent ee87938 commit 86dffa9f5486c688c61470b049fc7fb6c2d74f7e
Showing with 1 addition and 1 deletion.
  1. +1 −1 README.md
View
@@ -47,7 +47,7 @@ Overview of our model:
After this, it first trains the baseline model for joint object and part detection and then it trains our model with object appearance and class branches, initialized with the previously trained baseline model.
Finally, it trains Offset Net and merges it with the model that contains the object branches to obtain our final model.
- **Model:** Training the baseline model takes about 8h on a Titan X GPU. On the same GPU, the model with object appearance and class takes about 10h and Offset Net takes 5h. If you just want to use them you can download the pretrained models in the installation step above. Then run the demo to see the test results.
-- **Results:** If the program executes correctly, it will print the per-class results in average precision and their mean (mAP) for each of the 105 part classes in PASCAL-Part and 20 objects classes in PASCAL VOC. The baseline model achieves 22.0% mAP for parts (48.7% mAP for objects) on the validation set using no external training data nor bounding-box regression, whereas the model with object appearance and class achieves 25.9% mAP (49.9% mAP for objects).
+- **Results:** If the program executes correctly, it will print the per-class results in average precision and their mean (mAP) for each of the 105 part classes in PASCAL-Part and 20 objects classes in PASCAL VOC. The baseline model achieves 22.0% mAP for parts (48.7% mAP for objects) on the validation set using no external training data nor bounding-box regression, whereas the model with object appearance and class achieves 25.9% mAP (49.9% mAP for objects). Finally, our full model with relative location achieves 27.3% mAP.
- **Note:** The results vary due to the random order of images presented during training. To reproduce the above results we fix the initial seed of the random number generator.

0 comments on commit 86dffa9

Please sign in to comment.