def test_graph_interface( self ):
# exercise the interface
g = pgmlink.HypothesesGraph()
n1 = g.addNode(0)
n2 = g.addNode(5)
n3 = g.addNode(7)
a1 = g.addArc(n1, n2)
a2 = g.addArc(n1,n3)
self.assertEqual( pgmlink.countNodes(g), 3)
self.assertEqual( pgmlink.countArcs(g), 2)
self.assertEqual( g.earliest_timestep(), 0 )
self.assertEqual( g.latest_timestep(), 7 )
g.erase(a2)
g.erase(n3)
self.assertEqual( pgmlink.countNodes(g), 2)
self.assertEqual( pgmlink.countArcs(g), 1)
self.assertTrue( g.valid(n1) )
self.assertTrue( g.valid(n2) )
self.assertTrue( not g.valid(n3) )
self.assertTrue( g.valid(a1) )
self.assertTrue( not g.valid(a2) )
def test_graph_interface(self):
# exercise the interface
g = pgmlink.HypothesesGraph()
n1 = g.addNode(0)
n2 = g.addNode(5)
n3 = g.addNode(7)
a1 = g.addArc(n1, n2)
a2 = g.addArc(n1, n3)
self.assertEqual(pgmlink.countNodes(g), 3)
self.assertEqual(pgmlink.countArcs(g), 2)
self.assertEqual(g.earliest_timestep(), 0)
self.assertEqual(g.latest_timestep(), 7)
g.erase(a2)
g.erase(n3)
self.assertEqual(pgmlink.countNodes(g), 2)
self.assertEqual(pgmlink.countArcs(g), 1)
self.assertTrue(g.valid(n1))
self.assertTrue(g.valid(n2))
self.assertTrue(not g.valid(n3))
self.assertTrue(g.valid(a1))
self.assertTrue(not g.valid(a2))
obj_size[0] = options.avg_obj_size
else:
options.avg_obj_size = obj_size
# load traxelstore
ts, fs, ndim, t0, t1, probGenerator, transitionClassifier = loadTraxelstoreAndTransitionClassifier(
options, ilp_fn, time_range, shape)
# build hypotheses graph
hypotheses_graph, n_it, a_it, fov = getHypothesesGraphAndIterators(
options, shape, t0, t1, ts, probGenerator, transitionClassifier,
options.skipLinks, options.skipLinksBias)
if probGenerator is None:
import pgmlink
numElements = pgmlink.countNodes(hypotheses_graph) + pgmlink.countArcs(
hypotheses_graph)
# get the map of node -> list(traxel) or just traxel
if options.without_tracklets:
traxelMap = hypotheses_graph.getNodeTraxelMap()
else:
traxelMap = hypotheses_graph.getNodeTrackletMap()
maxNumObjects = int(options.max_num_objects)
margin = float(options.border_width)
def detectionProbabilityFunc(traxel):
return getDetectionFeatures(traxel, maxNumObjects + 1)
def transitionProbabilityFunc(srcTraxel, destTraxel):
def track_subgraphs(graph,
time_range,
timesteps_per_segment,
segment_overlap_timesteps,
conservation_tracking_parameter,
fov,
ilp_fn,
ts,
fs,
t0,
trans_classifier,
uncertaintyParam
):
"""
Experiment: track only subgraphs of the full hypotheses graph with some overlap,
and then stitch the results together using fusion moves.
"""
# define which segments we have
num_segments = int(np.ceil(float((time_range[1] - time_range[0])) / (timesteps_per_segment - segment_overlap_timesteps)))
segments = [(time_range[0] + i * (timesteps_per_segment - segment_overlap_timesteps),
(time_range[0] + (i + 1) * timesteps_per_segment - i * segment_overlap_timesteps))
for i in xrange(num_segments)]
tmap = graph.getNodeTraxelMap()
solutions = {}
arc_solutions = {}
div_solutions = {}
original_out_dir = options.out_dir
# track all segments individually
for i, segment in enumerate(segments):
print("************** Creating subgraph for timesteps in {}".format(segment))
# use special out-dir per window
options.out_dir = original_out_dir.rstrip('/') + '/window_' + str(i) + '/'
try:
os.makedirs(options.out_dir)
except:
pass
# create subgraph for this segment
node_mask = track.NodeMask(graph)
n_it = track.NodeIt(graph)
for n in n_it:
node_mask[n] = segment[0] <= tmap[n].Timestep < segment[1]
arc_mask = track.ArcMask(graph)
a_it = track.ArcIt(graph)
for a in a_it:
arc_mask[a] = tmap[graph.source(a)].Timestep >= segment[0] and tmap[graph.target(a)].Timestep < segment[1]
subgraph = track.HypothesesGraph()
track.copy_hypotheses_subgraph(graph, subgraph, node_mask, arc_mask)
subgraph_node_origin_map = subgraph.getNodeOriginReferenceMap()
subgraph_arc_origin_map = subgraph.getArcOriginReferenceMap()
subgraph.initLabelingMaps()
# fix variables in overlap
if i > 0:
sub_tmap = subgraph.getNodeTraxelMap()
n_it = track.NodeIt(subgraph)
for n in n_it:
if segment[0] == sub_tmap[n].Timestep:
origin_node = subgraph_node_origin_map[n]
origin_node_id = graph.id(origin_node)
subgraph.addAppearanceLabel(n, solutions[origin_node_id][-1])
print "fixing node ", origin_node_id, " which is ", subgraph.id(n), " in subgraph"
print("Subgraph has {} nodes and {} arcs".format(track.countNodes(subgraph), track.countArcs(subgraph)))
# create subgraph tracker
subgraph_tracker = track.ConsTracking(subgraph,
ts,
conservation_tracking_parameter,
uncertaintyParam,
fov,
bool(options.size_dependent_detection_prob),
options.avg_obj_size[0],
options.mnd,
options.division_threshold)
all_events = subgraph_tracker.track(conservation_tracking_parameter, bool(i > 0))
if len(options.raw_filename) > 0 and len(options.reranker_weight_file) > 0:
# run merger resolving and feature extraction, which also returns the score of each proposal
region_features = multitrack.getRegionFeatures(ndim)
scores = multitrack.runMergerResolving(options,
subgraph_tracker,
ts,
fs,
subgraph,
ilp_fn,
all_events,
fov,
region_features,
trans_classifier,
segment[0],
True)
best_sol_idx = int(np.argmax(np.array(scores)))
subgraph.set_solution(best_sol_idx)
print("====> selected solution {} in window {} <=====".format(best_sol_idx, i))
else:
subgraph.set_solution(0)
print("Done tracking subgraph")
# collect solutions
subgraph_node_active_map = subgraph.getNodeActiveMap()
subgraph_arc_active_map = subgraph.getArcActiveMap()
subgraph_div_active_map = subgraph.getDivisionActiveMap()
n_it = track.NodeIt(subgraph)
for n in n_it:
origin_node = subgraph_node_origin_map[n]
origin_node_id = graph.id(origin_node)
value = subgraph_node_active_map[n]
if not origin_node_id in solutions:
solutions[origin_node_id] = [value]
else:
solutions[origin_node_id].append(value)
div_solutions[origin_node_id] = subgraph_div_active_map[n]
a_it = track.ArcIt(subgraph)
for a in a_it:
origin_arc = subgraph_arc_origin_map[a]
origin_arc_id = graph.id(origin_arc)
arc_solutions[origin_arc_id] = subgraph_arc_active_map[a]
print("Done storing solutions")
# reset out-dir
options.out_dir = original_out_dir
# find overlapping variables
print("Computing overlap statistics...")
num_overlap_vars = sum([1 for values in solutions.values() if len(values) > 1])
num_disagreeing_overlap_vars = sum([1 for values in solutions.values() if len(values) > 1 and values[0] != values[1]])
for key, values in solutions.items():
if len(values) > 1 and values[0] != values[1]:
print("\tFound disagreement at {}: {} != {}".format(key, values[0], values[1]))
print("Found {} variables in overlaps, of which {} did disagree ({}%)".format(num_overlap_vars,
num_disagreeing_overlap_vars,
100.0 * float(num_disagreeing_overlap_vars) / num_overlap_vars))
if num_disagreeing_overlap_vars == 0:
# write overall solution back to hypotheses graph
graph.initLabelingMaps()
n_it = track.NodeIt(graph)
for n in n_it:
n_id = graph.id(n)
graph.addAppearanceLabel(n, solutions[n_id][-1])
graph.addDisappearanceLabel(n, solutions[n_id][-1])
# store division information
graph.addDivisionLabel(n, div_solutions[n_id])
# activate arcs
a_it = track.ArcIt(graph)
for a in a_it:
a_id = graph.id(a)
graph.addArcLabel(a, arc_solutions[a_id])
graph.set_injected_solution()
else:
raise AssertionError("Nodes did disagree, cannot create stitched solution")
if options.avg_obj_size != 0:
obj_size[0] = options.avg_obj_size
else:
options.avg_obj_size = obj_size
# load traxelstore
ts, fs, ndim, t0, t1, probGenerator, transitionClassifier = loadTraxelstoreAndTransitionClassifier(options, ilp_fn,
time_range,
shape)
# build hypotheses graph
hypotheses_graph, n_it, a_it, fov = getHypothesesGraphAndIterators(options, shape, t0, t1, ts, probGenerator, transitionClassifier, options.skipLinks, options.skipLinksBias)
if probGenerator is None:
import pgmlink
numElements = pgmlink.countNodes(hypotheses_graph) + pgmlink.countArcs(hypotheses_graph)
# get the map of node -> list(traxel) or just traxel
if options.without_tracklets:
traxelMap = hypotheses_graph.getNodeTraxelMap()
else:
traxelMap = hypotheses_graph.getNodeTrackletMap()
maxNumObjects = int(options.max_num_objects)
margin = float(options.border_width)
def detectionProbabilityFunc(traxel):
return getDetectionFeatures(traxel, maxNumObjects + 1)
def transitionProbabilityFunc(srcTraxel, destTraxel):
def track_subgraphs(graph,
time_range,
timesteps_per_segment,
segment_overlap_timesteps,
conservation_tracking_parameter,
fov,
ilp_fn,
ts,
fs,
t0,
trans_classifier,
uncertaintyParam
):
"""
Experiment: track only subgraphs of the full hypotheses graph with some overlap,
and then stitch the results together using fusion moves.
"""
# define which segments we have
num_segments = int(np.ceil(float((time_range[1] - time_range[0])) / (timesteps_per_segment - segment_overlap_timesteps)))
segments = [(time_range[0] + i * (timesteps_per_segment - segment_overlap_timesteps),
(time_range[0] + (i + 1) * timesteps_per_segment - i * segment_overlap_timesteps))
for i in xrange(num_segments)]
tmap = graph.getNodeTraxelMap()
solutions = {}
arc_solutions = {}
div_solutions = {}
original_out_dir = options.out_dir
# track all segments individually
for i, segment in enumerate(segments):
print("************** Creating subgraph for timesteps in {}".format(segment))
# use special out-dir per window
options.out_dir = original_out_dir.rstrip('/') + '/window_' + str(i) + '/'
try:
os.makedirs(options.out_dir)
except:
pass
# create subgraph for this segment
node_mask = track.NodeMask(graph)
n_it = track.NodeIt(graph)
for n in n_it:
node_mask[n] = segment[0] <= tmap[n].Timestep < segment[1]
arc_mask = track.ArcMask(graph)
a_it = track.ArcIt(graph)
for a in a_it:
arc_mask[a] = tmap[graph.source(a)].Timestep >= segment[0] and tmap[graph.target(a)].Timestep < segment[1]
subgraph = track.HypothesesGraph()
track.copy_hypotheses_subgraph(graph, subgraph, node_mask, arc_mask)
subgraph_node_origin_map = subgraph.getNodeOriginReferenceMap()
subgraph_arc_origin_map = subgraph.getArcOriginReferenceMap()
subgraph.initLabelingMaps()
# fix variables in overlap
if i > 0:
sub_tmap = subgraph.getNodeTraxelMap()
n_it = track.NodeIt(subgraph)
for n in n_it:
if segment[0] == sub_tmap[n].Timestep:
origin_node = subgraph_node_origin_map[n]
origin_node_id = graph.id(origin_node)
subgraph.addAppearanceLabel(n, solutions[origin_node_id][-1])
print "fixing node ", origin_node_id, " which is ", subgraph.id(n), " in subgraph"
print("Subgraph has {} nodes and {} arcs".format(track.countNodes(subgraph), track.countArcs(subgraph)))
# create subgraph tracker
subgraph_tracker = track.ConsTracking(subgraph,
ts,
conservation_tracking_parameter,
uncertaintyParam,
fov,
bool(options.size_dependent_detection_prob),
options.avg_obj_size[0],
options.mnd,
options.division_threshold)
all_events = subgraph_tracker.track(conservation_tracking_parameter, bool(i > 0))
if len(options.raw_filename) > 0 and len(options.reranker_weight_file) > 0:
# run merger resolving and feature extraction, which also returns the score of each proposal
region_features = multitrack.getRegionFeatures(ndim)
scores = multitrack.runMergerResolving(options,
subgraph_tracker,
ts,
fs,
subgraph,
ilp_fn,
all_events,
fov,
region_features,
trans_classifier,
segment[0],
True)
best_sol_idx = int(np.argmax(np.array(scores)))
subgraph.set_solution(best_sol_idx)
print("====> selected solution {} in window {} <=====".format(best_sol_idx, i))
else:
subgraph.set_solution(0)
print("Done tracking subgraph")
# collect solutions
subgraph_node_active_map = subgraph.getNodeActiveMap()
subgraph_arc_active_map = subgraph.getArcActiveMap()
subgraph_div_active_map = subgraph.getDivisionActiveMap()
n_it = track.NodeIt(subgraph)
for n in n_it:
origin_node = subgraph_node_origin_map[n]
origin_node_id = graph.id(origin_node)
value = subgraph_node_active_map[n]
if not origin_node_id in solutions:
solutions[origin_node_id] = [value]
else:
solutions[origin_node_id].append(value)
div_solutions[origin_node_id] = subgraph_div_active_map[n]
a_it = track.ArcIt(subgraph)
for a in a_it:
origin_arc = subgraph_arc_origin_map[a]
origin_arc_id = graph.id(origin_arc)
arc_solutions[origin_arc_id] = subgraph_arc_active_map[a]
print("Done storing solutions")
# reset out-dir
options.out_dir = original_out_dir
# find overlapping variables
print("Computing overlap statistics...")
num_overlap_vars = sum([1 for values in solutions.values() if len(values) > 1])
num_disagreeing_overlap_vars = sum([1 for values in solutions.values() if len(values) > 1 and values[0] != values[1]])
for key, values in solutions.iteritems():
if len(values) > 1 and values[0] != values[1]:
print("\tFound disagreement at {}: {} != {}".format(key, values[0], values[1]))
print("Found {} variables in overlaps, of which {} did disagree ({}%)".format(num_overlap_vars,
num_disagreeing_overlap_vars,
100.0 * float(num_disagreeing_overlap_vars) / num_overlap_vars))
if num_disagreeing_overlap_vars == 0:
# write overall solution back to hypotheses graph
graph.initLabelingMaps()
n_it = track.NodeIt(graph)
for n in n_it:
n_id = graph.id(n)
graph.addAppearanceLabel(n, solutions[n_id][-1])
graph.addDisappearanceLabel(n, solutions[n_id][-1])
# store division information
graph.addDivisionLabel(n, div_solutions[n_id])
# activate arcs
a_it = track.ArcIt(graph)
for a in a_it:
a_id = graph.id(a)
graph.addArcLabel(a, arc_solutions[a_id])
graph.set_injected_solution()
else:
raise AssertionError("Nodes did disagree, cannot create stitched solution")