shapenet / jointembedding Goto Github PK

Hi, when I tried to reproduce the image-based shape retrieval experiment,
I got a very low accuracy: top100-acc only reached about 10%, but in your paper (fig.8), top100-acc>95% on the benchmark(105 shapes and 315 images) . The dataset I used is shapenet-02691156 (airplanes), and the corresponding synthetic images with overlaid background, each shape is associated with 50 images.
Is the low accuracy on Shapenet-02691156 normal? Looking forward to your reply.

I run the pipeline and it succeed until step 3 run_shape_embedding_training.sh. The python script render_lfd_shape_list.py throws a FileNotFoundError for shape_embedding/lfd_images. I think shape_embedding folder is missing in data folder. But if I create the folder myself, the rendering jobs of render_lfd_shape_list.py all failed. Can you give me some tips to fix it?

I followed the 5 steps on the README file, ran the shell script run_joint_embedding_training.sh, which took only a few minutes to finish. However I still cannot test any images. Either the script prepare_testing.py or image_based_shape_retrieval.py in /src/image_embedding_testing/ were not able to run because file not found: /JointEmbedding/data/image_embedding/image_embedding_training_03001627_rcnn/snapshots_03001627_iter_20000.caffemodel
I think I must miss something, but I wonder what that is.

python extract_image_embedding.py --image E:\JointEmbedding\src\image_embeddin g_testing\testing_images --caffemodel image_embedding_03001627.caffemodel --pro totxt image_embedding_03001627.prototxt 运行这个脚本，但是出现错误Segmentation fault
我的电脑配置：Windows7 ，内存4G，64位，处理器：AMD Athlon（tm）II X2 240Processor 2.8GHz
等待您的回复，谢谢

The shape training computes HOG from views with a size of 120x120, 60x60 and 30x30. However, how does it come possible since HOG cells have a size of 8x8 which is not a multiple of 60 nor 30.

I tried the live demo ,selected the picture and region and submited,but the server seemed to refuse service

In the file utilites_math.py, there exist two two camPosToQuaternion functions. I get the same input to these two functions and they return different results.
`def camPosToQuaternion(cx, cy, cz):
camDist = math.sqrt(cx * cx + cy * cy + cz * cz)
cx = cx / camDist
cy = cy / camDist
cz = cz / camDist
axis = (-cz, 0, cx)
angle = math.acos(cy)
a = math.sqrt(2) / 2
b = math.sqrt(2) / 2
w1 = axis[0]
w2 = axis[1]
w3 = axis[2]
c = math.cos(angle / 2)
d = math.sin(angle / 2)
q1 = a * c - b * d * w1
q2 = b * c + a * d * w1
q3 = a * d * w2 + b * d * w3
q4 = -b * d * w2 + a * d * w3
return (q1, q2, q3, q4)

def quaternionFromYawPitchRoll(yaw, pitch, roll):
c1 = math.cos(yaw / 2.0)
c2 = math.cos(pitch / 2.0)
c3 = math.cos(roll / 2.0)
s1 = math.sin(yaw / 2.0)
s2 = math.sin(pitch / 2.0)
s3 = math.sin(roll / 2.0)
q1 = c1 * c2 * c3 + s1 * s2 * s3
q2 = c1 * c2 * s3 - s1 * s2 * c3
q3 = c1 * s2 * c3 + s1 * c2 * s3
q4 = s1 * c2 * c3 - c1 * s2 * s3
return (q1, q2, q3, q4)

def camPosToQuaternion(cx, cy, cz):
q1a = 0
q1b = 0
q1c = math.sqrt(2) / 2
q1d = math.sqrt(2) / 2
camDist = math.sqrt(cx * cx + cy * cy + cz * cz)
cx = cx / camDist
cy = cy / camDist
cz = cz / camDist
t = math.sqrt(cx * cx + cy * cy)
tx = cx / t
ty = cy / t
yaw = math.acos(ty)
if tx > 0:
yaw = 2 * math.pi - yaw
pitch = 0
tmp = min(max(txcx + tycy, -1),1)
#roll = math.acos(tx * cx + ty * cy)
roll = math.acos(tmp)
if cz < 0:
roll = -roll
print("%f %f %f" % (yaw, pitch, roll))
q2a, q2b, q2c, q2d = quaternionFromYawPitchRoll(yaw, pitch, roll)
q1 = q1a * q2a - q1b * q2b - q1c * q2c - q1d * q2d
q2 = q1b * q2a + q1a * q2b + q1d * q2c - q1c * q2d
q3 = q1c * q2a - q1d * q2b + q1a * q2c + q1b * q2d
q4 = q1d * q2a + q1c * q2b - q1b * q2c + q1a * q2d
return (q1, q2, q3, q4)
`

These code snippets are in JointEmbedding\src \siamese_embedding\prepare_training.py.

When I debug them, the result is not what we except.

So I modified "model_id_to_image_ids = [[]]*shape_num" to "model_id_to_image_ids = [[] for i in range(shape_num)]". Then the result is right.