varun-suresh/clustering

Implements "Clustering a Million Faces by Identity"

clustering's Issues

Understanding feature extraction

I have been trying to understand the feature extraction process. My approach to this is from the book creative illustration Beautiful paper 😃

how to set the value of n_neighbors and thresh? when most of face in datasets is isolated, I think your method will be invalid?

cluster(vectors, n_neighbors=200, thresh= [1.1])
what is your principle of setting the n_neighbors and thresh?

moreover, when most of face in datasets is isolated, I think your method will be invalid?

Can i ask how to produce the vectors of new data like the form of lfw? i saw the .mat is complicated .

Can i ask how to produce the vectors of new data like the form of lfw? i saw the .mat is complicated .

running error

Hi Varun,
I test your code,and it appeared:
~/Clustering$ python demo.py --lfw_path ./lfw -v LightenedCNN_A_lfw.mat

Distance calculation time : 52.4087600708
N Clusters: 6854, thresh: 1.1
No of clusters: 6854
Threshold : 1.1
Traceback (most recent call last):
File "demo.py", line 113, in
f1_score = evaluate_clusters(clusters['clusters'], labels_lookup)
File "demo.py", line 64, in evaluate_clusters
precision, recall = calculate_pairwise_pr(clusters, labels_lookup)
File "/home/king/Clustering/evaluation.py", line 30, in calculate_pairwise_pr
cp, tp = count_correct_pairs(cluster, labels_lookup)
File "/home/king/Clustering/evaluation.py", line 16, in count_correct_pairs
if labels_lookup[f1] == labels_lookup[f2]:
KeyError: 8944

But when i tried B and C,it is ok,here is the result:

~/Clustering$ python demo.py --lfw_path ./lfw -v LightenedCNN_C_lfw.mat
Distance calculation time : 51.666009903
N Clusters: 6987, thresh: 1.1
No of clusters: 6987
Threshold : 1.1
Correct Pairs that are in the same cluster:196778
Total pairs as per the clusters created: 204500
Total possible true pairs:242225.0
Precision : 0.962239608802
Recall : 0.812376922283
f1_score : 0.880980468969

And another question is about the visualization,
when i tried json.load(),it got:

File "visualize.py", line 72, in
clusters = json.load(args.clusters)
File "/usr/lib/python2.7/json/init.py", line 286, in load
return loads(fp.read(),
AttributeError: 'str' object has no attribute 'read'

when i tried json.loads(),it got:

File "visualize.py", line 72, in
clusters = json.loads(args.clusters)
File "/usr/lib/python2.7/json/init.py", line 338, in loads
return _default_decoder.decode(s)
File "/usr/lib/python2.7/json/decoder.py", line 366, in decode
obj, end = self.raw_decode(s, idx=_w(s, 0).end())
File "/usr/lib/python2.7/json/decoder.py", line 384, in raw_decode
raise ValueError("No JSON object could be decoded")

please give me some tips,thank you~~~

get different results as you provided

[Security] Path Traversal Vulnerability found

A path traversal attack (also known as directory traversal) aims to access files and directories that are stored outside the web root folder. By manipulating variables that reference files with “dot-dot-slash (../)” sequences and its variations or by using absolute file paths, it may be possible to access arbitrary files and directories stored on file system including application source code or configuration and critical system files. It should be noted that access to files is limited by system operational access control (such as in the case of locked or in-use files on the Microsoft Windows operating system).

This attack is also known as “dot-dot-slash”, “directory traversal”, “directory climbing” and “backtracking”.

Root Cause Analysis

In this case, the path traversal vulnerability can be blamed on incorrect usage of the send_from_directory Flask call. The vulnerability occurs due to the code snippet shown below

Clustering/visualize.py

Lines 33 to 36 in 53e663e

	@app.route('/img/<path:fpath>', methods=['GET'])
	def get_img_path(fpath):
	print(os.path.dirname(fpath), fpath.split('/')[-1])
	return send_from_directory(os.path.dirname(fpath),fpath.split('/')[-1])

Here, since the fpath parameter is attacker controlled, the effective directory and filename passed to the send_from_directory call can be controlled by the attacker leading to a path traversal attack.

Proof of Concept

The bug can be verified using the proof of concept similar to the one shown below.

curl -i --path-as-is -s -k -X $'GET' \
    -H $'Host: 0.0.0.0:5000' -H $'User-Agent: Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:91.0) Gecko/20100101 Firefox/91.0' -H $'Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,*/*;q=0.8' -H $'Accept-Language: en-US,en;q=0.5' -H $'Accept-Encoding: gzip, deflate' -H $'Connection: close' -H $'Upgrade-Insecure-Requests: 1' \
    $'http://0.0.0.0:5000/img/../../../../../../etc/passwd'

Remediation

This can be easily fixed my restricting the value of file and path parameters by a fixed whitelist of possible values.

CVSS 3 Score

CVSS:3.0/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H

This bug was found using CodeQL by Github

New data clustering

Does it support new data clustering based on the model which has already converged?

Precision drop?

Hi Varun,

were you able to resolve, why your implementation has such a huge drop in precision? Thanks and looking forward to your reply.

best, Vivek

Usage

Could this work on a dir of unlabeled faces?

run error

When I cluster.py
I got this error ::
Traceback (most recent call last):
File "clustering.py", line 4, in
import pyflann
File "/Users/youyin/anaconda3/lib/python3.6/site-packages/pyflann/init.py", line 27, in
from index import *
ModuleNotFoundError: No module named 'index'

My python version is 3.6.4
Do you have any suggestions?Thank you