rebuttal.tex

We thank reviewers for their time and valuable comments. We will
update the paper as outlined below.

Why L2 normalization works (R18,R34,R38)
The proposed re-normalization procedure can be interpreted as a simple
form of affine calibration where the uncalibrated SVM score s(x) = w^T
x + b is transformed into a calibrated score f(x) = alpha * s(x) +
beta, where alpha = 1/||w|| and beta = -b*alpha. Note that the bias
term b is not ignored but used to compute f(x). The intuition is that
for L2 normalized descriptors x computing the calibrated score f(x)
corresponds to computing the normalized cross-correlation between
vectors w and x. This was found to work well in retrieval (see e.g.
Sivic and Zisserman, ICCV 2003 [32]) or matching whitened HOG
descriptors [11,13] (see e.g. [Doersch et al., Mid-Level Visual
Element Discovery as discriminative Mode Seeking, NIPS 2013]), but has
not yet been used in conjunction with SVMs as done in this work. Note
also that the affine calibration can be computed offline and does not
need any additional computation or storage at query time compared to
the p-value calibration of Gronat et al. [12].

-> Section 3.2

Comparison with p-value calibration of Gronat et al. [12] (R18,R34,R38)
We have compared the proposed method with the p-value calibration of
(Gronat et al.[12]) applied to Fisher vectors. The results show that
our method significantly outperforms the p-value calibration (37.8%
recall@1 vs. 32.3% recall@1; 56.9% recall@5 vs. 50.0% recall@5) for
the Fisher vector of dimension 128. Similar gains are observed for
other FV dimensions.

-> Now in tables. Please discuss in "Comparison to other methods"

Applying the proposed method to other descriptors (R18, R38)
The proposed approach can be applied to other descriptors beyond
Fisher vectors. To demonstrate this we have applied the proposed
calibration by re-normalization to the bag-of-visual-words descriptor
used in (Gronat et al.[12]) and observed improvement from 29.4%
recall@K=1 (vanilla bag-of-visual-words) to 32.6% recall@K=1 (e-SVM
with the proposed calibration by renormalization). For
bag-of-visual-words, the e-SVM with p-value calibration gives
comparable results (33.6% recall@K=1) but at an increased training
cost and a significantly higher required memory footprint at query
time, which makes the method very hard to scale-up to database sizes
beyond 200k images.

-> Create new paragraph in the result section "Applying the proposed
method to other descriptors".
Refer to table.


Significance
- The exemplar SVM has been recently used for a number of recognition
task. This paper describes a simple procedure to avoid calibration of
the exemplar SVM score at query time by re-normalizing the learnt SVM
weights.

- We demonstrate that proposed method is applicable to two different
descriptors, the bag-of-visual-words and Fisher vectors.
-> Conclusion

- We show the method performs on par or better than other competing
calibration methods for the task of large-scale place recognition.

- We believe the paper is significant and could be adopted by others
because the method is simple, scalable, does not require additional
storage or computation at query time and could be applied for other
tasks that use exemplar SVM beyond the addressed place recognition
application.




Rejecting query images that are out of the database (R18)
We agree with R18 that the method cannot reject query images that are
out of the database. Similar to other retrieval methods, we address a
ranking problem and not a classification one. Given a query image the
goal is to create a ranked list of N best candidate images. These
candidates can be further verified by another more expensive method
(e.g. geometric verification [Philbin et al., Object retrieval with
large vocabularies and fast spatial matching, CVPR 2007, Knopp et al.,
Avoiding confusing features in place recognition, ECCV 2010 ]) to
reject false positives or to reject a query that comes from an unseen
place.


Other comments:
- Correction to figure 4 (R34)
Reviewer R34 is right, both the second (1st row) and the third (2nd
row) retrieved images for the second query example should be marked as
wrong. We will correct the figure for the next revision.
done

- Clarification of comparison to other works, L313-315 (R34)
We compare our method to the results from [12] who re-implemented [17]
and reported the results on the Pittsburgh 25k dataset.
results are in tables.

- Notation (R34)
We will improve the notation by showing vectors as bolds following the
suggestion from R34.
-> todo?

- Lower performance for high dimensional FV (R34)
It is known that dimensionality reduction can have a positive effect
on the FV performance, where lower dimensional descriptors can
outperform higher dimensional descriptors. This is a known effect also
observed in previous work, see e.g. Table 1 in [15].
- add comment to result section with a citaiton.

- Details of the bag-of-visual-words experiment (R38)
We use the same setup as [12] except that our descriptors are
rootSIFTs instead of SURFs used in [12].

-> Make sure we mention what do.