filtered_cache_images.hdf5 training not working

[Rob174]

Context

Dataset with patches

• not in margins
• always with seep and/or spill annotation
• augmentation factor of 100

Problem

➡️ Model converges

But unsatisfactory result

Prediction

Compared to the reference

➡️ 2 problems:

• Background (= class other = with no annotation) : is always confused with the seep class
• Spill : is not recognized

Diagnosis

Statistics of number of classes present on patches

• spill_only: 50 patches
• seep_only: 13248 patches
• seep_spill: 9828 patches

Compared with original polygons statistics

To get them we have used

1. the DB Manager to execute SQL queries to get a. the number of rasters with each amount of seep possible b. the number of rasters with each amount of spill possible
2. a python script to merge the two tables (gotten with copy paste) and count the number of raster with each amount of seep and spill possible

Number of seep	Number of spill	Number of raster		Number of seep	Number of spill	Number of raster
0	0	357	2	0	4
0	1	10	2	1	1
0	2	11	2	5	1
0	3	10	2	8	1
0	4	5	20	0	2
0	5	2	3	0	9
0	6	3	3	1	1
0	7	1	3	3	1
0	9	1	4	0	9
1	0	11	4	3	1
1	3	2	5	0	6
1	6	1	5	1	2
1	7	1	5	2	1
10	0	2	6	0	3
11	0	3	6	2	2
11	4	1	7	0	5
12	0	1	8	0	3
13	0	1	8	1	1
14	7	1	8	2	1
15	0	2	8	7	1
15	1	1	9	0	2

(with 0,0 point excluded) Interactive vizualization

[Rob174]

Conclusions : cause of the problem

• Patch with no annotations

not given to the model ➡️ understandable that the model does not well on these patches

• Spill class misclassified

According to Statistics of number of classes present on patches, there are less patches with spill maybe due to the fact that several spill polygons can be on the same image

But according to Compared with original polygons statistics, there are 196 spill and 533 seep on all rasters

➡️this issue will be treated in issue 32

[Rob174]

Second cause of problems

Interpolation on labels cause class shifts:

On this image at the origin there was probably only the brightest class of value 2, but with interpolation (due to augmentations), the class 1 has been added

➡️ This problem will be treated in issue 31

[Rob174]

Only 4 interpolation methods allowed for warpAffine

[Rob174]

Algo 1 : do not work

We observe on an example that we can bound number of intermediate pixels with artificial values generated around real value pixels.

Potential solution: Step 1: warpAffine Step 2: convert to uint8 Step 3: maxpooling with a kernel of more than the number of pixels

(kernel 5) (kernel 10) (kernel 3)

[Rob174]

Algo 2:

1. Gaussian blur
2. Subtract it from the original image

Filtered because delete all labels

[Rob174]

Slow solution but add precision

Create the augmented patch directly with points of annotations

Done in 3636648

[Rob174]

After computing another time the classes stats on the new filtered cache we observe a more balanced dataset regarding seep and spill classes repartition:	seep_only	spill_only	seep_spill
11766	9258	208
55 %	44 %	0.98 %